Practice Exams:

The Open Group OG0-021 Dumps

EXAM: ArchiMate 2 Part 1
Verified By Experts
Premium Bundle
$49.99

OG0-021 Premium ETE File


  • Premium File 80 Questions & Answers
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Free OG0-021 Exam Questions in ETE Format

The open group.pass4sure.og0-021.v2026-10-30.by.jaxon.48q.ete
Downloads: 25
Size: 2.22 MB

Passing the The Open Group OG0-021 ArchiMate 2 Part 1 Exam has never been easier. Latest The Open Group OG0-021 Exam Dumps with Accurate Questions & Answers to prepare for the exam. ETE Files for The Open Group OG0-021 ArchiMate 2 Part 1 practice test questions are uploaded by users who have taken the exams in last few weeks. Vumingo is the new generation ETE Exam Simulator Software which is used to open these ETE files which can be converted to PDF format as well. The Open Group OG0-021 Premium Exam Questions, Video Training Courses & Study Guide is also available which has been verified by PrepAway Experts. Be one of the success stories and pass the ArchiMate 2 Part 1 Exam and share your comments with the community.

Last Week Results!

10

Customers Passed The Open Group OG0-021 Exam

88%

Average Score In The Exam At Testing Centre

83%

Questions came word for word from this dump

ArchiMate 2 Part 1 Exams Facts
IAPP AIGP | 800 Original Unofficial Educational Practice Questions

Question 1

A city benefits agency uses a hand-written eligibility rule beside a model that predicts missed appointments. Which step would provide the most targeted control for the problem described?

  1. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  2. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  3. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  4. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

Explanation

Correct Answer

A:

Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic. The practical acceptance evidence is: document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Incorrect Answers

B:

The alternative—classify a text-generating model separately from a numerical risk predictor when preparing evaluations—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

C:

The alternative—assess confidentiality, data minimization and adversarial misuse together in the AI risk register—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

D:

The alternative—define triggers, reporting channels, containment authority and post-incident learning for AI harms—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Question 2

A team reviewing the following situation asks for the most relevant governance principle: A city benefits agency uses a hand-written eligibility rule beside a model that predicts missed appointments. Which principle best explains the required control?

  1. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  2. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  3. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  4. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Explanation

Correct Answer

B:

Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic. That is why the case calls for this action: distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.

Incorrect Answers

A:

That explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

C:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

D:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Question 3

The organization has implemented a control for this case: A city benefits agency uses a hand-written eligibility rule beside a model that predicts missed appointments. Which result would best validate it?

  1. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  2. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  4. Use role-based exercises and verify that participants can identify triggers for escalation and human review.

Explanation

Correct Answer

C:

Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Incorrect Answers

A:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

B:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

D:

Use role-based exercises and verify that participants can identify triggers for escalation and human review would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Question 4

A team considers this AI use: A city benefits agency uses a hand-written eligibility rule beside a model that predicts missed appointments. Rather than applying the relevant control, it will label every software rule as machine learning and apply a single undifferentiated test checklist. What can go wrong?

  1. Variable outputs may include failures never observed in the single demonstration.
  2. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  3. Changes can bypass controls even when new risks appear in training data or deployment.
  4. The inventory hides different failure mechanisms and assigns ineffective testing to both components.

Explanation

Correct Answer

D:

The inventory hides different failure mechanisms and assigns ineffective testing to both components. The shortcut overlooks the central safeguard: distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Incorrect Answers

A:

The stated consequence—variable outputs may include failures never observed in the single demonstration—belongs to a different failure mode. In this case, the inventory hides different failure mechanisms and assigns ineffective testing to both components.

B:

The stated consequence—unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest—belongs to a different failure mode. In this case, the inventory hides different failure mechanisms and assigns ineffective testing to both components.

C:

The stated consequence—changes can bypass controls even when new risks appear in training data or deployment—belongs to a different failure mode. In this case, the inventory hides different failure mechanisms and assigns ineffective testing to both components.

Question 5

A city benefits agency uses a hand-written eligibility rule beside a model that predicts missed appointments. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  2. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  3. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  4. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Explanation

Correct Answer

A:

Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component. Together, they implement and test the intended safeguard. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Incorrect Answers

B:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

C:

This pair supports another governance intervention: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

D:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Question 6

A freight cooperative compares a fixed tax calculator with a trained demand-forecasting service. Which step would provide the most targeted control for the problem described?

  1. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.
  2. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  3. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  4. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

Explanation

Correct Answer

B:

For the decision, distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic. The practical acceptance evidence is: document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component

Incorrect Answers

A:

The alternative—compare relevant error outcomes across affected groups and investigate material disparities before deployment—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

C:

For the decision, the alternative—classify a text-generating model separately from a numerical risk predictor when preparing evaluations—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

D:

For the decision, the alternative—define triggers, reporting channels, containment authority and post-incident learning for AI harms—would not resolve the stated issue. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

Question 7

A team reviewing the following situation asks for the most relevant governance principle: A freight cooperative compares a fixed tax calculator with a trained demand-forecasting service. Which principle best explains the required control?

  1. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  2. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  3. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

C:

For the technical explanation, models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic. That is why the case calls for this action: distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls

Incorrect Answers

A:

For the technical explanation, that explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

B:

For the technical explanation, that explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

D:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.

Question 8

The organization has implemented a control for this case: A freight cooperative compares a fixed tax calculator with a trained demand-forecasting service. Which result would best validate it?

  1. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  2. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  4. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Explanation

Correct Answer

D:

For the verification, document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

Incorrect Answers

A:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

B:

For the verification, review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component

C:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Question 9

A freight cooperative compares a fixed tax calculator with a trained demand-forecasting service. The team instead proposes: Label every software rule as machine learning and apply a single undifferentiated test checklist. Which failure should the reviewer expect most directly?

  1. The inventory hides different failure mechanisms and assigns ineffective testing to both components.
  2. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  3. Unsupported narratives can pass release even when a separate numeric score appears accurate.
  4. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

Explanation

Correct Answer

A:

For the risk comparison, the inventory hides different failure mechanisms and assigns ineffective testing to both components. The shortcut overlooks the central safeguard: distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

Incorrect Answers

B:

For the risk comparison, the stated consequence—unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest—belongs to a different failure mode. In this case, the inventory hides different failure mechanisms and assigns ineffective testing to both components

C:

The stated consequence—unsupported narratives can pass release even when a separate numeric score appears accurate—belongs to a different failure mode. In this case, the inventory hides different failure mechanisms and assigns ineffective testing to both components.

D:

This consequence follows more directly from another shortcut: Let every involved team assume another team owns the final risk decision. The shortcut actually described would instead create this risk: The inventory hides different failure mechanisms and assigns ineffective testing to both components.

Question 10

Choose the intervention and validation that fit this business problem: A freight cooperative compares a fixed tax calculator with a trained demand-forecasting service.

  1. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  2. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  3. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.

Explanation

Correct Answer

B:

For the action and verification, distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component. Together, they implement and test the intended safeguard. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic

Incorrect Answers

A:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

C:

This pair supports another governance intervention: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

D:

This pair supports another governance intervention: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For the actual case, the required action is: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. The supporting evidence is: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Question 11

An AI project team faces this situation: An insurance help desk uses a claim-priority score and a separate model to draft customer explanations. What should the team do first to manage the central issue?

  1. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  2. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.
  3. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  4. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Explanation

Correct Answer

C:

Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures. The practical acceptance evidence is: review samples of generated text for supported claims and separately test predictor error on held-out cases.

Incorrect Answers

A:

The alternative—assess confidentiality, data minimization and adversarial misuse together in the AI risk register—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

B:

The alternative—set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

D:

The alternative—establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Question 12

An insurance help desk uses a claim-priority score and a separate model to draft customer explanations. Why is the recommended governance approach appropriate?

  1. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.
  2. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  3. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  4. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Explanation

Correct Answer

D:

Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures. That is why the case calls for this action: classify a text-generating model separately from a numerical risk predictor when preparing evaluations.

Incorrect Answers

A:

That explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

B:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

C:

That explains why a team might train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Question 13

An insurance help desk uses a claim-priority score and a separate model to draft customer explanations. Which recorded test or observation most directly demonstrates an effective response?

  1. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  2. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  3. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  4. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Explanation

Correct Answer

A:

Review samples of generated text for supported claims and separately test predictor error on held-out cases. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Incorrect Answers

B:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases.

C:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases.

D:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases.

Question 14

An insurance help desk uses a claim-priority score and a separate model to draft customer explanations. The team instead proposes: Use only the predictor's numerical accuracy measure to approve both the prediction and generated explanation. Which failure should the reviewer expect most directly?

  1. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  2. Unsupported narratives can pass release even when a separate numeric score appears accurate.
  3. Variable outputs may include failures never observed in the single demonstration.
  4. The organization cannot substantiate who approved a change or what evidence existed.

Explanation

Correct Answer

B:

Unsupported narratives can pass release even when a separate numeric score appears accurate. The shortcut overlooks the central safeguard: classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate.

C:

This consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate.

D:

This consequence follows more directly from another shortcut: Keep all AI risk approvals in informal chats with no linked version history. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate.

Question 15

The governance team needs both a decision and a way to verify it for this case: An insurance help desk uses a claim-priority score and a separate model to draft customer explanations. Which pair is best?

  1. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  2. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  3. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Explanation

Correct Answer

C:

Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Review samples of generated text for supported claims and separately test predictor error on held-out cases. Together, they implement and test the intended safeguard. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Incorrect Answers

A:

This pair supports another governance intervention: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

B:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

D:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

Question 16

A museum uses a visitor-volume forecast alongside a system that writes exhibit descriptions. Which governance action most directly addresses the need in this situation?

  1. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.
  2. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  3. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  4. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.

Explanation

Correct Answer

D:

For the decision, classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures. The practical acceptance evidence is: review samples of generated text for supported claims and separately test predictor error on held-out cases

Incorrect Answers

A:

The alternative—compare relevant error outcomes across affected groups and investigate material disparities before deployment—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

B:

The alternative—distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

C:

The alternative—publish an AI governance responsibility matrix naming accountable risk owners and approval authorities—would not resolve the stated issue. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Question 17

A museum uses a visitor-volume forecast alongside a system that writes exhibit descriptions. Which explanation best connects the risk to the proposed governance action?

  1. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  2. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  3. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

A:

For the technical explanation, generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures. That is why the case calls for this action: classify a text-generating model separately from a numerical risk predictor when preparing evaluations

Incorrect Answers

B:

That explains why a team might distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

C:

For the technical explanation, that explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures

D:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.

Question 18

An AI governance reviewer examines this situation: A museum uses a visitor-volume forecast alongside a system that writes exhibit descriptions. Which acceptance check is most decisive?

  1. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  2. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Explanation

Correct Answer

B:

For the verification, review samples of generated text for supported claims and separately test predictor error on held-out cases. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures

Incorrect Answers

A:

For the verification, run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases

C:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases.

D:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to review samples of generated text for supported claims and separately test predictor error on held-out cases.

Question 19

A museum uses a visitor-volume forecast alongside a system that writes exhibit descriptions. The team instead proposes: Use only the predictor's numerical accuracy measure to approve both the prediction and generated explanation. Which failure should the reviewer expect most directly?

  1. A well-performing average can conceal severe errors concentrated in one affected group.
  2. Variable outputs may include failures never observed in the single demonstration.
  3. Unsupported narratives can pass release even when a separate numeric score appears accurate.
  4. Ongoing harm may continue because nobody has authority or a clear reporting route.

Explanation

Correct Answer

C:

For the risk comparison, unsupported narratives can pass release even when a separate numeric score appears accurate. The shortcut overlooks the central safeguard: classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate

D:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Unsupported narratives can pass release even when a separate numeric score appears accurate.

Question 20

The governance team needs both a decision and a way to verify it for this case: A museum uses a visitor-volume forecast alongside a system that writes exhibit descriptions. Which pair is best?

  1. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  2. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  3. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  4. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

Explanation

Correct Answer

D:

For the action and verification, classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Review samples of generated text for supported claims and separately test predictor error on held-out cases. Together, they implement and test the intended safeguard. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures

Incorrect Answers

A:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

B:

For the action and verification, this pair supports another governance intervention: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases

C:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. The supporting evidence is: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

Question 21

A university deploys a guidance assistant that answers the same admissions question with different phrasing. Which step would provide the most targeted control for the problem described?

  1. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.
  2. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  3. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  4. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Explanation

Correct Answer

A:

Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria. The practical acceptance evidence is: run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Incorrect Answers

B:

The alternative—distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

C:

The alternative—assess confidentiality, data minimization and adversarial misuse together in the AI risk register—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

D:

The alternative—apply tiered governance requirements based on impact, organizational capacity and risk tolerance—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Question 22

A team reviewing the following situation asks for the most relevant governance principle: A university deploys a guidance assistant that answers the same admissions question with different phrasing. Which principle best explains the required control?

  1. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  2. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.
  3. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  4. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Explanation

Correct Answer

B:

Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria. That is why the case calls for this action: set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.

Incorrect Answers

A:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

C:

That explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

D:

That explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Question 23

The organization has implemented a control for this case: A university deploys a guidance assistant that answers the same admissions question with different phrasing. Which result would best validate it?

  1. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  2. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  4. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Explanation

Correct Answer

C:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Incorrect Answers

A:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

B:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

D:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Question 24

A university deploys a guidance assistant that answers the same admissions question with different phrasing. The team instead proposes: Treat one satisfactory model response as proof that all future responses will be equally safe. Which failure should the reviewer expect most directly?

  1. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  2. The inventory hides different failure mechanisms and assigns ineffective testing to both components.
  3. Ongoing harm may continue because nobody has authority or a clear reporting route.
  4. Variable outputs may include failures never observed in the single demonstration.

Explanation

Correct Answer

D:

Variable outputs may include failures never observed in the single demonstration. The shortcut overlooks the central safeguard: set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration.

B:

This consequence follows more directly from another shortcut: Label every software rule as machine learning and apply a single undifferentiated test checklist. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration.

C:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration.

Question 25

Choose the intervention and validation that fit this business problem: A university deploys a guidance assistant that answers the same admissions question with different phrasing.

  1. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  2. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  3. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Explanation

Correct Answer

A:

Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance. Together, they implement and test the intended safeguard. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Incorrect Answers

B:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

C:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

D:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Question 26

An AI project team faces this situation: A legal intake team observes slightly different summaries when the same document is submitted twice. What should the team do first to manage the central issue?

  1. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  2. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.
  3. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  4. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Explanation

Correct Answer

B:

For the decision, set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria. The practical acceptance evidence is: run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance

Incorrect Answers

A:

For the decision, the alternative—distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria

C:

The alternative—classify a text-generating model separately from a numerical risk predictor when preparing evaluations—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

D:

For the decision, the alternative—apply tiered governance requirements based on impact, organizational capacity and risk tolerance—would not resolve the stated issue. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria

Question 27

A team reviewing the following situation asks for the most relevant governance principle: A legal intake team observes slightly different summaries when the same document is submitted twice. Which principle best explains the required control?

  1. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  2. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.
  3. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

C:

For the technical explanation, some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria. That is why the case calls for this action: set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run

Incorrect Answers

A:

That explains why a team might distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

B:

That explains why a team might compare relevant error outcomes across affected groups and investigate material disparities before deployment, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

D:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Question 28

An AI governance reviewer examines this situation: A legal intake team observes slightly different summaries when the same document is submitted twice. Which acceptance check is most decisive?

  1. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  2. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  3. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  4. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

D:

For the verification, run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria

Incorrect Answers

A:

For the verification, review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance

B:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

C:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Question 29

The organization has this problem: A legal intake team observes slightly different summaries when the same document is submitted twice. It takes a shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. What is the principal risk?

  1. Variable outputs may include failures never observed in the single demonstration.
  2. The inventory hides different failure mechanisms and assigns ineffective testing to both components.
  3. Unsupported narratives can pass release even when a separate numeric score appears accurate.
  4. Unknown subcontractors, data rights and change risks can reach production unchecked.

Explanation

Correct Answer

A:

For the risk comparison, variable outputs may include failures never observed in the single demonstration. The shortcut overlooks the central safeguard: set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Label every software rule as machine learning and apply a single undifferentiated test checklist. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration

C:

This consequence follows more directly from another shortcut: Use only the predictor's numerical accuracy measure to approve both the prediction and generated explanation. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration.

D:

This consequence follows more directly from another shortcut: Accept a vendor marketing statement as the only evidence of its AI risk controls. The shortcut actually described would instead create this risk: Variable outputs may include failures never observed in the single demonstration.

Question 30

Choose the intervention and validation that fit this business problem: A legal intake team observes slightly different summaries when the same document is submitted twice.

  1. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  2. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  3. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Explanation

Correct Answer

B:

For the action and verification, set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance. Together, they implement and test the intended safeguard. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria

Incorrect Answers

A:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

C:

This pair supports another governance intervention: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

D:

This pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. The supporting evidence is: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Question 31

Consider the following proposed AI use: A housing nonprofit sees high average screening accuracy but unusually high rejection errors for one applicant group. Which response best addresses the governance gap?

  1. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  2. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  3. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.
  4. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

Explanation

Correct Answer

C:

Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The practical acceptance evidence is: review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Incorrect Answers

A:

This would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

B:

This would help address a separate governance task: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

D:

This would help address a separate governance task: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

Question 32

A housing nonprofit sees high average screening accuracy but unusually high rejection errors for one applicant group. Which explanation best connects the risk to the proposed governance action?

  1. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  2. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  3. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  4. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Explanation

Correct Answer

D:

Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. That is why the case calls for this action: compare relevant error outcomes across affected groups and investigate material disparities before deployment.

Incorrect Answers

A:

That explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

B:

That explains why a team might distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

C:

That explains why a team might attach explicit review and approval conditions to design, test, release and monitoring milestones, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Question 33

A housing nonprofit sees high average screening accuracy but unusually high rejection errors for one applicant group. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  2. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  3. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Use role-based exercises and verify that participants can identify triggers for escalation and human review.

Explanation

Correct Answer

A:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Incorrect Answers

B:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions.

C:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions.

D:

Use role-based exercises and verify that participants can identify triggers for escalation and human review would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Question 34

For this case—A housing nonprofit sees high average screening accuracy but unusually high rejection errors for one applicant group—staff choose the shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. Which consequence best explains why this is inadequate?

  1. Variable outputs may include failures never observed in the single demonstration.
  2. A well-performing average can conceal severe errors concentrated in one affected group.
  3. The inventory hides different failure mechanisms and assigns ineffective testing to both components.
  4. A harmful or unsuitable use case may become expensive to correct after development.

Explanation

Correct Answer

B:

A well-performing average can conceal severe errors concentrated in one affected group. The shortcut overlooks the central safeguard: compare relevant error outcomes across affected groups and investigate material disparities before deployment. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group.

C:

This consequence follows more directly from another shortcut: Label every software rule as machine learning and apply a single undifferentiated test checklist. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group.

D:

This consequence follows more directly from another shortcut: Allow teams to build a model first and identify its intended users only at launch. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group.

Question 35

Choose the intervention and validation that fit this business problem: A housing nonprofit sees high average screening accuracy but unusually high rejection errors for one applicant group.

  1. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  2. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

C:

Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Review group-level false-positive and false-negative rates with population context and document mitigation decisions. Together, they implement and test the intended safeguard. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Incorrect Answers

A:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

B:

This pair supports another governance intervention: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

D:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Question 36

Consider the following proposed AI use: A lending cooperative finds a strong overall credit score while certain geographic communities receive more false rejections. Which response best addresses the governance gap?

  1. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  2. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  3. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.
  4. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

Explanation

Correct Answer

D:

For the decision, compare relevant error outcomes across affected groups and investigate material disparities before deployment. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The practical acceptance evidence is: review group-level false-positive and false-negative rates with population context and document mitigation decisions

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment

B:

This would help address a separate governance task: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

C:

This would help address a separate governance task: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For this case, however, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. The more direct step is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

Question 37

The organization chooses to compare relevant error outcomes across affected groups and investigate material disparities before deployment for this situation: A lending cooperative finds a strong overall credit score while certain geographic communities receive more false rejections. Which rationale best supports that choice?

  1. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.
  2. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  3. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  4. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Explanation

Correct Answer

A:

For the technical explanation, aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute. That is why the case calls for this action: compare relevant error outcomes across affected groups and investigate material disparities before deployment

Incorrect Answers

B:

For the technical explanation, that explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute

C:

For the technical explanation, that explains why a team might distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute

D:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Question 38

The organization has implemented a control for this case: A lending cooperative finds a strong overall credit score while certain geographic communities receive more false rejections. Which result would best validate it?

  1. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  2. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  3. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  4. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.

Explanation

Correct Answer

B:

For the verification, review group-level false-positive and false-negative rates with population context and document mitigation decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute

Incorrect Answers

A:

For the verification, test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions

C:

Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions.

D:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Question 39

The organization has this problem: A lending cooperative finds a strong overall credit score while certain geographic communities receive more false rejections. It takes a shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. What is the principal risk?

  1. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  2. Variable outputs may include failures never observed in the single demonstration.
  3. A well-performing average can conceal severe errors concentrated in one affected group.
  4. The team may miss legal, societal and practical harms outside its own expertise.

Explanation

Correct Answer

C:

For the risk comparison, a well-performing average can conceal severe errors concentrated in one affected group. The shortcut overlooks the central safeguard: compare relevant error outcomes across affected groups and investigate material disparities before deployment. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group

D:

This consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: A well-performing average can conceal severe errors concentrated in one affected group.

Question 40

The governance team needs both a decision and a way to verify it for this case: A lending cooperative finds a strong overall credit score while certain geographic communities receive more false rejections. Which pair is best?

  1. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  2. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  3. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Explanation

Correct Answer

D:

For the action and verification, compare relevant error outcomes across affected groups and investigate material disparities before deployment. Review group-level false-positive and false-negative rates with population context and document mitigation decisions. Together, they implement and test the intended safeguard. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute

Incorrect Answers

A:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

B:

For the action and verification, this pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions

C:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. The supporting evidence is: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Question 41

An AI project team faces this situation: A clinic experiments with a medical summarizer that may expose patient identifiers in responses. What should the team do first to manage the central issue?

  1. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  2. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  3. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.
  4. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Explanation

Correct Answer

A:

Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Training and inference can expose personal data or permit misuse even when the model predicts accurately. The practical acceptance evidence is: test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Incorrect Answers

B:

This would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, training and inference can expose personal data or permit misuse even when the model predicts accurately. The more direct step is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

C:

The alternative—set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run—would not resolve the stated issue. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

D:

This would help address a separate governance task: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For this case, however, training and inference can expose personal data or permit misuse even when the model predicts accurately. The more direct step is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

Question 42

A clinic experiments with a medical summarizer that may expose patient identifiers in responses. Why is the recommended governance approach appropriate?

  1. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.
  2. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  3. Models that learn patterns from data need validation of training effects, while explicit rules can be traced directly through their prescribed logic.
  4. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Explanation

Correct Answer

B:

Training and inference can expose personal data or permit misuse even when the model predicts accurately. That is why the case calls for this action: assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

Incorrect Answers

A:

That explains why a team might compare relevant error outcomes across affected groups and investigate material disparities before deployment, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

C:

That explains why a team might distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

D:

That explains why a team might map the vendor's model responsibilities separately from the organization's deployment and use controls, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Question 43

The organization has implemented a control for this case: A clinic experiments with a medical summarizer that may expose patient identifiers in responses. Which result would best validate it?

  1. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  2. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  3. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  4. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Explanation

Correct Answer

C:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Incorrect Answers

A:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

B:

Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

D:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Question 44

The organization has this problem: A clinic experiments with a medical summarizer that may expose patient identifiers in responses. It takes a shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. What is the principal risk?

  1. Variable outputs may include failures never observed in the single demonstration.
  2. Unsupported narratives can pass release even when a separate numeric score appears accurate.
  3. Changes can bypass controls even when new risks appear in training data or deployment.
  4. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

Explanation

Correct Answer

D:

Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest. The shortcut overlooks the central safeguard: assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

B:

This consequence follows more directly from another shortcut: Use only the predictor's numerical accuracy measure to approve both the prediction and generated explanation. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

C:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

Question 45

A clinic experiments with a medical summarizer that may expose patient identifiers in responses. Which action and evidence pair most directly addresses the situation?

  1. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  2. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Explanation

Correct Answer

A:

Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries. Together, they implement and test the intended safeguard. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Incorrect Answers

B:

This pair supports another governance intervention: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

C:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

D:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Question 46

Consider the following proposed AI use: A human-resources team uses a resume assistant that retains more applicant details than needed. Which response best addresses the governance gap?

  1. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.
  2. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  3. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  4. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

Explanation

Correct Answer

B:

For the decision, assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Training and inference can expose personal data or permit misuse even when the model predicts accurately. The practical acceptance evidence is: test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries

Incorrect Answers

A:

This would help address a separate governance task: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For this case, however, training and inference can expose personal data or permit misuse even when the model predicts accurately. The more direct step is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

C:

This would help address a separate governance task: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For this case, however, training and inference can expose personal data or permit misuse even when the model predicts accurately. The more direct step is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

D:

This would help address a separate governance task: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For this case, however, training and inference can expose personal data or permit misuse even when the model predicts accurately. The more direct step is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

Question 47

A team reviewing the following situation asks for the most relevant governance principle: A human-resources team uses a resume assistant that retains more applicant details than needed. Which principle best explains the required control?

  1. Generative output requires assessment of factual grounding and unsafe content in addition to task performance; prediction requires appropriate error and calibration measures.
  2. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.
  3. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  4. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Explanation

Correct Answer

C:

For the technical explanation, training and inference can expose personal data or permit misuse even when the model predicts accurately. That is why the case calls for this action: assess confidentiality, data minimization and adversarial misuse together in the AI risk register

Incorrect Answers

A:

That explains why a team might classify a text-generating model separately from a numerical risk predictor when preparing evaluations, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

B:

That explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

D:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Question 48

A human-resources team uses a resume assistant that retains more applicant details than needed. Which recorded test or observation most directly demonstrates an effective response?

  1. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  2. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  3. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

D:

For the verification, test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Training and inference can expose personal data or permit misuse even when the model predicts accurately

Incorrect Answers

A:

For the verification, document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries

B:

For the verification, review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries

C:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Question 49

For this case—A human-resources team uses a resume assistant that retains more applicant details than needed—staff choose the shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. Which consequence best explains why this is inadequate?

  1. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  2. A well-performing average can conceal severe errors concentrated in one affected group.
  3. Variable outputs may include failures never observed in the single demonstration.
  4. Unknown subcontractors, data rights and change risks can reach production unchecked.

Explanation

Correct Answer

A:

For the risk comparison, unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest. The shortcut overlooks the central safeguard: assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Training and inference can expose personal data or permit misuse even when the model predicts accurately

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest

D:

This consequence follows more directly from another shortcut: Accept a vendor marketing statement as the only evidence of its AI risk controls. The shortcut actually described would instead create this risk: Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

Question 50

Choose the intervention and validation that fit this business problem: A human-resources team uses a resume assistant that retains more applicant details than needed.

  1. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  2. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  3. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  4. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Explanation

Correct Answer

B:

For the action and verification, assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries. Together, they implement and test the intended safeguard. Training and inference can expose personal data or permit misuse even when the model predicts accurately

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries

C:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

D:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. The supporting evidence is: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Question 51

A regional bank has product, compliance and data science teams but nobody signs off AI risk decisions. Which step would provide the most targeted control for the problem described?

  1. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.
  2. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  3. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  4. Attach explicit review and approval conditions to design, test, release and monitoring milestones.

Explanation

Correct Answer

C:

Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Shared participation does not eliminate the need for named accountability for launch, controls and incidents. The practical acceptance evidence is: inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Incorrect Answers

A:

This would help address a separate governance task: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

B:

This would help address a separate governance task: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

D:

This would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Question 52

A team reviewing the following situation asks for the most relevant governance principle: A regional bank has product, compliance and data science teams but nobody signs off AI risk decisions. Which principle best explains the required control?

  1. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  2. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  3. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  4. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Explanation

Correct Answer

D:

Shared participation does not eliminate the need for named accountability for launch, controls and incidents. That is why the case calls for this action: publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Incorrect Answers

A:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

B:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

C:

That explains why a team might define triggers, reporting channels, containment authority and post-incident learning for AI harms, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Question 53

The organization has implemented a control for this case: A regional bank has product, compliance and data science teams but nobody signs off AI risk decisions. Which result would best validate it?

  1. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  2. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Check meeting records for documented challenges, responses, dissent and resulting design changes.
  4. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Explanation

Correct Answer

A:

Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Incorrect Answers

B:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

C:

Check meeting records for documented challenges, responses, dissent and resulting design changes would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

D:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Question 54

A regional bank has product, compliance and data science teams but nobody signs off AI risk decisions. The team instead proposes: Let every involved team assume another team owns the final risk decision. Which failure should the reviewer expect most directly?

  1. The team may miss legal, societal and practical harms outside its own expertise.
  2. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.
  3. Resources are wasted on low-risk work while significant risks may still go unidentified.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

B:

Material risks can remain unresolved because no person is accountable for accepting or rejecting them. The shortcut overlooks the central safeguard: publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

C:

This consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

D:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

Question 55

A regional bank has product, compliance and data science teams but nobody signs off AI risk decisions. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  2. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  3. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  4. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Explanation

Correct Answer

C:

Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage. Together, they implement and test the intended safeguard. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Incorrect Answers

A:

This pair supports another governance intervention: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

B:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

D:

This pair supports another governance intervention: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Question 56

A government service contractor has five project sponsors and no named owner for its AI release gate. Which governance action most directly addresses the need in this situation?

  1. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  2. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  3. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.
  4. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Explanation

Correct Answer

D:

For the decision, publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Shared participation does not eliminate the need for named accountability for launch, controls and incidents. The practical acceptance evidence is: inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage

Incorrect Answers

A:

This would help address a separate governance task: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

B:

This would help address a separate governance task: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

C:

This would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, shared participation does not eliminate the need for named accountability for launch, controls and incidents. The more direct step is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Question 57

The organization chooses to publish an AI governance responsibility matrix naming accountable risk owners and approval authorities for this situation: A government service contractor has five project sponsors and no named owner for its AI release gate. Which rationale best supports that choice?

  1. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  2. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  3. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

A:

For the technical explanation, shared participation does not eliminate the need for named accountability for launch, controls and incidents. That is why the case calls for this action: publish an AI governance responsibility matrix naming accountable risk owners and approval authorities

Incorrect Answers

B:

That explains why a team might train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

C:

That explains why a team might apply tiered governance requirements based on impact, organizational capacity and risk tolerance, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

D:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.

Question 58

An AI governance reviewer examines this situation: A government service contractor has five project sponsors and no named owner for its AI release gate. Which acceptance check is most decisive?

  1. Check meeting records for documented challenges, responses, dissent and resulting design changes.
  2. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  3. Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Explanation

Correct Answer

B:

For the verification, inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Shared participation does not eliminate the need for named accountability for launch, controls and incidents

Incorrect Answers

A:

For the verification, check meeting records for documented challenges, responses, dissent and resulting design changes would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage

C:

Use role-based exercises and verify that participants can identify triggers for escalation and human review would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

D:

Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance would test a different outcome. The directly relevant evidence is to inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Question 59

A government service contractor has five project sponsors and no named owner for its AI release gate. The team instead proposes: Let every involved team assume another team owns the final risk decision. Which failure should the reviewer expect most directly?

  1. The team may miss legal, societal and practical harms outside its own expertise.
  2. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  3. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

C:

For the risk comparison, material risks can remain unresolved because no person is accountable for accepting or rejecting them. The shortcut overlooks the central safeguard: publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Shared participation does not eliminate the need for named accountability for launch, controls and incidents

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them

B:

This consequence follows more directly from another shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Material risks can remain unresolved because no person is accountable for accepting or rejecting them

Question 60

The governance team needs both a decision and a way to verify it for this case: A government service contractor has five project sponsors and no named owner for its AI release gate. Which pair is best?

  1. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  3. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.
  4. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Explanation

Correct Answer

D:

For the action and verification, publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage. Together, they implement and test the intended safeguard. Shared participation does not eliminate the need for named accountability for launch, controls and incidents

Incorrect Answers

A:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

B:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

C:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. The supporting evidence is: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.

Question 61

Consider the following proposed AI use: A university plans a disciplinary-risk model without involving student support or legal staff. Which response best addresses the governance gap?

  1. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.
  2. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.
  3. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  4. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Explanation

Correct Answer

A:

Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The practical acceptance evidence is: check meeting records for documented challenges, responses, dissent and resulting design changes.

Incorrect Answers

B:

This would help address a separate governance task: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

C:

This would help address a separate governance task: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

D:

This would help address a separate governance task: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

Question 62

The organization chooses to convene legal, security, domain experts and affected-user representatives before accepting model design decisions for this situation: A university plans a disciplinary-risk model without involving student support or legal staff. Which rationale best supports that choice?

  1. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  2. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  3. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  4. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Explanation

Correct Answer

B:

Different disciplines reveal failure modes that a purely technical or purely legal review may overlook. That is why the case calls for this action: convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

Incorrect Answers

A:

That explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

C:

That explains why a team might apply tiered governance requirements based on impact, organizational capacity and risk tolerance, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

D:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Question 63

A university plans a disciplinary-risk model without involving student support or legal staff. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  2. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Check meeting records for documented challenges, responses, dissent and resulting design changes.
  4. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

C:

Check meeting records for documented challenges, responses, dissent and resulting design changes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Incorrect Answers

A:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes.

B:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes.

D:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes.

Question 64

For this case—A university plans a disciplinary-risk model without involving student support or legal staff—staff choose the shortcut: Allow only the model builders to review a system that will affect diverse external users. Which consequence best explains why this is inadequate?

  1. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  2. Resources are wasted on low-risk work while significant risks may still go unidentified.
  3. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  4. The team may miss legal, societal and practical harms outside its own expertise.

Explanation

Correct Answer

D:

The team may miss legal, societal and practical harms outside its own expertise. The shortcut overlooks the central safeguard: convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise.

B:

This consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise.

C:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise.

Question 65

The governance team needs both a decision and a way to verify it for this case: A university plans a disciplinary-risk model without involving student support or legal staff. Which pair is best?

  1. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  2. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

A:

Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Check meeting records for documented challenges, responses, dissent and resulting design changes. Together, they implement and test the intended safeguard. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Incorrect Answers

B:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

C:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

D:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

Question 66

Consider the following proposed AI use: A hospital considers a triage assistant using only the vendor's technical evaluation team. Which response best addresses the governance gap?

  1. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.
  2. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.
  3. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  4. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Explanation

Correct Answer

B:

For the decision, convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The practical acceptance evidence is: check meeting records for documented challenges, responses, dissent and resulting design changes

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions

C:

This would help address a separate governance task: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

D:

This would help address a separate governance task: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. For this case, however, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. The more direct step is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

Question 67

A hospital considers a triage assistant using only the vendor's technical evaluation team. Why is the recommended governance approach appropriate?

  1. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  2. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.
  3. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  4. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Explanation

Correct Answer

C:

For the technical explanation, different disciplines reveal failure modes that a purely technical or purely legal review may overlook. That is why the case calls for this action: convene legal, security, domain experts and affected-user representatives before accepting model design decisions

Incorrect Answers

A:

For the technical explanation, that explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook

B:

That explains why a team might map the vendor's model responsibilities separately from the organization's deployment and use controls, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

D:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Question 68

The organization has implemented a control for this case: A hospital considers a triage assistant using only the vendor's technical evaluation team. Which result would best validate it?

  1. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  2. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Check meeting records for documented challenges, responses, dissent and resulting design changes.

Explanation

Correct Answer

D:

For the verification, check meeting records for documented challenges, responses, dissent and resulting design changes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook

Incorrect Answers

A:

For the verification, verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes

B:

For the verification, review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes

C:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to check meeting records for documented challenges, responses, dissent and resulting design changes.

Question 69

A hospital considers a triage assistant using only the vendor's technical evaluation team. The team instead proposes: Allow only the model builders to review a system that will affect diverse external users. Which failure should the reviewer expect most directly?

  1. The team may miss legal, societal and practical harms outside its own expertise.
  2. Resources are wasted on low-risk work while significant risks may still go unidentified.
  3. Leadership cannot see whether risks are rising or critical issues remain open.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

A:

For the risk comparison, the team may miss legal, societal and practical harms outside its own expertise. The shortcut overlooks the central safeguard: convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise

C:

This consequence follows more directly from another shortcut: Report only the number of AI pilots completed as evidence of responsible governance. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise.

D:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: The team may miss legal, societal and practical harms outside its own expertise.

Question 70

The governance team needs both a decision and a way to verify it for this case: A hospital considers a triage assistant using only the vendor's technical evaluation team. Which pair is best?

  1. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  2. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  3. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  4. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Explanation

Correct Answer

B:

For the action and verification, convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Check meeting records for documented challenges, responses, dissent and resulting design changes. Together, they implement and test the intended safeguard. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook

Incorrect Answers

A:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

C:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

D:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. The supporting evidence is: Check meeting records for documented challenges, responses, dissent and resulting design changes.

Question 71

A call-center operator is asked to handle an unreliable assistant but receives only a general AI video. Which step would provide the most targeted control for the problem described?

  1. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  2. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  3. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.
  4. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Explanation

Correct Answer

C:

Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Incorrect Answers

A:

This would help address a separate governance task: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For this case, however, a generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The more direct step is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

B:

This would help address a separate governance task: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For this case, however, a generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The more direct step is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

D:

The alternative—record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 72

The organization chooses to train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control for this situation: A call-center operator is asked to handle an unreliable assistant but receives only a general AI video. Which rationale best supports that choice?

  1. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.
  2. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  3. Training and inference can expose personal data or permit misuse even when the model predicts accurately.
  4. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

Explanation

Correct Answer

D:

A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. That is why the case calls for this action: train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

Incorrect Answers

A:

That explains why a team might map the vendor's model responsibilities separately from the organization's deployment and use controls, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

B:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

C:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

Question 73

The organization has implemented a control for this case: A call-center operator is asked to handle an unreliable assistant but receives only a general AI video. Which result would best validate it?

  1. Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  2. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Explanation

Correct Answer

A:

Use role-based exercises and verify that participants can identify triggers for escalation and human review. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

Incorrect Answers

B:

The alternative—review board or executive reports for risk trends, exceptions, decisions and follow-up actions—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

C:

The alternative—verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

D:

The alternative—review group-level false-positive and false-negative rates with population context and document mitigation decisions—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 74

The organization has this problem: A call-center operator is asked to handle an unreliable assistant but receives only a general AI video. It takes a shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. What is the principal risk?

  1. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.
  2. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  3. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

B:

Staff may not know how to respond to unsafe outputs or governance events in their own workflow. The shortcut overlooks the central safeguard: train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Let every involved team assume another team owns the final risk decision. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

C:

This consequence follows more directly from another shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

D:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

Question 75

A call-center operator is asked to handle an unreliable assistant but receives only a general AI video. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  2. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  3. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

C:

Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Use role-based exercises and verify that participants can identify triggers for escalation and human review. Together, they implement and test the intended safeguard. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

Incorrect Answers

A:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. The supporting evidence is: Use role-based exercises and verify that participants can identify triggers for escalation and human review.

B:

This pair supports another governance intervention: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For the actual case, the required action is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. The supporting evidence is: Use role-based exercises and verify that participants can identify triggers for escalation and human review.

D:

The alternative—set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 76

A manufacturing manager must approve model changes after attending a generic introductory webinar. Which governance action most directly addresses the need in this situation?

  1. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  2. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  3. Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.
  4. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

Explanation

Correct Answer

D:

For the decision, train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review

Incorrect Answers

A:

This would help address a separate governance task: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For this case, however, a generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The more direct step is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.

B:

For the decision, this would help address a separate governance task: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For this case, however, a generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. The more direct step is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control

C:

The alternative—require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 77

A team reviewing the following situation asks for the most relevant governance principle: A manufacturing manager must approve model changes after attending a generic introductory webinar. Which principle best explains the required control?

  1. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  2. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  3. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  4. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Explanation

Correct Answer

A:

For the technical explanation, a generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties. That is why the case calls for this action: train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control

Incorrect Answers

B:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

C:

That explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.

D:

For the technical explanation, that explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties

Question 78

A manufacturing manager must approve model changes after attending a generic introductory webinar. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  2. Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  3. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  4. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Explanation

Correct Answer

B:

For the verification, use role-based exercises and verify that participants can identify triggers for escalation and human review. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties

Incorrect Answers

A:

The alternative—inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

C:

The alternative—compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

D:

The alternative—check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance—would not resolve the stated issue. The practical acceptance evidence is: use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 79

For this case—A manufacturing manager must approve model changes after attending a generic introductory webinar—staff choose the shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. Which consequence best explains why this is inadequate?

  1. Resources are wasted on low-risk work while significant risks may still go unidentified.
  2. Leadership cannot see whether risks are rising or critical issues remain open.
  3. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  4. Employees receive inconsistent instructions for data ingestion, retention and disclosure.

Explanation

Correct Answer

C:

For the risk comparison, staff may not know how to respond to unsafe outputs or governance events in their own workflow. The shortcut overlooks the central safeguard: train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

B:

This consequence follows more directly from another shortcut: Report only the number of AI pilots completed as evidence of responsible governance. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

D:

This consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

Question 80

Choose the intervention and validation that fit this business problem: A manufacturing manager must approve model changes after attending a generic introductory webinar.

  1. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  2. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  3. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.
  4. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.

Explanation

Correct Answer

D:

For the action and verification, train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Use role-based exercises and verify that participants can identify triggers for escalation and human review. Together, they implement and test the intended safeguard. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For the actual case, the required action is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. The supporting evidence is: Use role-based exercises and verify that participants can identify triggers for escalation and human review

B:

This pair supports another governance intervention: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For the actual case, the required action is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. The supporting evidence is: Use role-based exercises and verify that participants can identify triggers for escalation and human review.

C:

This pair supports another governance intervention: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For the actual case, the required action is: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. The supporting evidence is: Use role-based exercises and verify that participants can identify triggers for escalation and human review.

Question 81

A small charity pilots a meeting-note helper and separately plans a benefits eligibility model. Which governance action most directly addresses the need in this situation?

  1. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  2. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  3. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  4. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.

Explanation

Correct Answer

A:

Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The practical acceptance evidence is: compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Incorrect Answers

B:

This would help address a separate governance task: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

C:

This would help address a separate governance task: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

D:

This would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Question 82

A team reviewing the following situation asks for the most relevant governance principle: A small charity pilots a meeting-note helper and separately plans a benefits eligibility model. Which principle best explains the required control?

  1. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  2. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  3. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

B:

A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. That is why the case calls for this action: apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Incorrect Answers

A:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

C:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

D:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Question 83

The organization has implemented a control for this case: A small charity pilots a meeting-note helper and separately plans a benefits eligibility model. Which result would best validate it?

  1. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  3. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  4. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Explanation

Correct Answer

C:

Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Incorrect Answers

A:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

B:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

D:

Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Question 84

A small charity pilots a meeting-note helper and separately plans a benefits eligibility model. The team instead proposes: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. Which failure should the reviewer expect most directly?

  1. The team may miss legal, societal and practical harms outside its own expertise.
  2. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  3. Variable outputs may include failures never observed in the single demonstration.
  4. Resources are wasted on low-risk work while significant risks may still go unidentified.

Explanation

Correct Answer

D:

Resources are wasted on low-risk work while significant risks may still go unidentified. The shortcut overlooks the central safeguard: apply tiered governance requirements based on impact, organizational capacity and risk tolerance. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified.

B:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified.

C:

This consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified.

Question 85

A small charity pilots a meeting-note helper and separately plans a benefits eligibility model. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  2. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  3. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  4. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Explanation

Correct Answer

A:

Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions. Together, they implement and test the intended safeguard. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Incorrect Answers

B:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

C:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

D:

This pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Question 86

Consider the following proposed AI use: A multinational retailer uses an internal text classifier and an automated hiring-shortlist system. Which response best addresses the governance gap?

  1. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.
  2. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  3. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  4. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

Explanation

Correct Answer

B:

For the decision, apply tiered governance requirements based on impact, organizational capacity and risk tolerance. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The practical acceptance evidence is: compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions

Incorrect Answers

A:

This would help address a separate governance task: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

C:

For the decision, this would help address a separate governance task: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance

D:

This would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. The more direct step is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Question 87

A multinational retailer uses an internal text classifier and an automated hiring-shortlist system. Why is the recommended governance approach appropriate?

  1. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  2. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  3. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  4. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Explanation

Correct Answer

C:

For the technical explanation, a low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight. That is why the case calls for this action: apply tiered governance requirements based on impact, organizational capacity and risk tolerance

Incorrect Answers

A:

That explains why a team might train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

B:

For the technical explanation, that explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight

D:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Question 88

The organization has implemented a control for this case: A multinational retailer uses an internal text classifier and an automated hiring-shortlist system. Which result would best validate it?

  1. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  2. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  4. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Explanation

Correct Answer

D:

For the verification, compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight

Incorrect Answers

A:

For the verification, verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions

B:

For the verification, review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions

C:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Question 89

The organization has this problem: A multinational retailer uses an internal text classifier and an automated hiring-shortlist system. It takes a shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. What is the principal risk?

  1. Resources are wasted on low-risk work while significant risks may still go unidentified.
  2. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  3. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  4. Ongoing harm may continue because nobody has authority or a clear reporting route.

Explanation

Correct Answer

A:

For the risk comparison, resources are wasted on low-risk work while significant risks may still go unidentified. The shortcut overlooks the central safeguard: apply tiered governance requirements based on impact, organizational capacity and risk tolerance. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified

D:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Resources are wasted on low-risk work while significant risks may still go unidentified.

Question 90

The governance team needs both a decision and a way to verify it for this case: A multinational retailer uses an internal text classifier and an automated hiring-shortlist system. Which pair is best?

  1. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  3. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

B:

For the action and verification, apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions. Together, they implement and test the intended safeguard. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight

Incorrect Answers

A:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

C:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

D:

This pair supports another governance intervention: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For the actual case, the required action is: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. The supporting evidence is: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Question 91

An AI project team faces this situation: A city authority buys a third-party chatbot but configures its own citizen advice workflows. What should the team do first to manage the central issue?

  1. Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control.
  2. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  3. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  4. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.

Explanation

Correct Answer

C:

Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The practical acceptance evidence is: verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Incorrect Answers

A:

The alternative—train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control—is not the most direct response to this case. The matching action is to map the vendor's model responsibilities separately from the organization's deployment and use controls.

B:

This would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The more direct step is: Map the vendor's model responsibilities separately from the organization's deployment and use controls.

D:

This would help address a separate governance task: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For this case, however, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The more direct step is: Map the vendor's model responsibilities separately from the organization's deployment and use controls.

Question 92

A city authority buys a third-party chatbot but configures its own citizen advice workflows. Why is the recommended governance approach appropriate?

  1. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  2. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  3. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  4. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Explanation

Correct Answer

D:

The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. That is why the case calls for this action: map the vendor's model responsibilities separately from the organization's deployment and use controls.

Incorrect Answers

A:

That explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

B:

That explains why a team might train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

C:

That explains why a team might attach explicit review and approval conditions to design, test, release and monitoring milestones, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Question 93

A city authority buys a third-party chatbot but configures its own citizen advice workflows. Which recorded test or observation most directly demonstrates an effective response?

  1. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  2. Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  3. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  4. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Explanation

Correct Answer

A:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Incorrect Answers

B:

Use role-based exercises and verify that participants can identify triggers for escalation and human review would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

C:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

D:

Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Question 94

For this case—A city authority buys a third-party chatbot but configures its own citizen advice workflows—staff choose the shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. Which consequence best explains why this is inadequate?

  1. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  2. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  3. The team may miss legal, societal and practical harms outside its own expertise.
  4. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.

Explanation

Correct Answer

B:

Local misuse, weak access control or poor monitoring can remain unmanaged by either party. The shortcut overlooks the central safeguard: map the vendor's model responsibilities separately from the organization's deployment and use controls. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party.

C:

This consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party.

D:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party.

Question 95

Choose the intervention and validation that fit this business problem: A city authority buys a third-party chatbot but configures its own citizen advice workflows.

  1. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  3. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Action: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. Evidence: Review samples of generated text for supported claims and separately test predictor error on held-out cases.

Explanation

Correct Answer

C:

Map the vendor's model responsibilities separately from the organization's deployment and use controls. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions. Together, they implement and test the intended safeguard. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Incorrect Answers

A:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

B:

This pair supports another governance intervention: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

D:

This pair supports another governance intervention: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Question 96

Consider the following proposed AI use: A retail chain licenses a forecasting API and chooses the stores and decisions that rely on it. Which response best addresses the governance gap?

  1. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  2. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.
  3. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  4. Map the vendor's model responsibilities separately from the organization's deployment and use controls.

Explanation

Correct Answer

D:

For the decision, map the vendor's model responsibilities separately from the organization's deployment and use controls. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The practical acceptance evidence is: verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The more direct step is: Map the vendor's model responsibilities separately from the organization's deployment and use controls

B:

This would help address a separate governance task: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For this case, however, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The more direct step is: Map the vendor's model responsibilities separately from the organization's deployment and use controls.

C:

This would help address a separate governance task: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For this case, however, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. The more direct step is: Map the vendor's model responsibilities separately from the organization's deployment and use controls.

Question 97

A retail chain licenses a forecasting API and chooses the stores and decisions that rely on it. Which explanation best connects the risk to the proposed governance action?

  1. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.
  2. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  3. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  4. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Explanation

Correct Answer

A:

For the technical explanation, the supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring. That is why the case calls for this action: map the vendor's model responsibilities separately from the organization's deployment and use controls

Incorrect Answers

B:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

C:

That explains why a team might apply tiered governance requirements based on impact, organizational capacity and risk tolerance, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

D:

That explains why a team might compare relevant error outcomes across affected groups and investigate material disparities before deployment, but it does not justify the decision required by the case. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Question 98

The organization has implemented a control for this case: A retail chain licenses a forecasting API and chooses the stores and decisions that rely on it. Which result would best validate it?

  1. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  2. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  3. Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Explanation

Correct Answer

B:

For the verification, verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring

Incorrect Answers

A:

Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

C:

For the verification, use role-based exercises and verify that participants can identify triggers for escalation and human review would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions

D:

Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component would test a different outcome. The directly relevant evidence is to verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Question 99

A retail chain licenses a forecasting API and chooses the stores and decisions that rely on it. The team instead proposes: Assume purchasing an AI service transfers every operational governance duty to the supplier. Which failure should the reviewer expect most directly?

  1. The team may miss legal, societal and practical harms outside its own expertise.
  2. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  3. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

C:

For the risk comparison, local misuse, weak access control or poor monitoring can remain unmanaged by either party. The shortcut overlooks the central safeguard: map the vendor's model responsibilities separately from the organization's deployment and use controls. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party

B:

For the risk comparison, this consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party

D:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Local misuse, weak access control or poor monitoring can remain unmanaged by either party.

Question 100

Choose the intervention and validation that fit this business problem: A retail chain licenses a forecasting API and chooses the stores and decisions that rely on it.

  1. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  3. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Explanation

Correct Answer

D:

For the action and verification, map the vendor's model responsibilities separately from the organization's deployment and use controls. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions. Together, they implement and test the intended safeguard. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions

B:

For the action and verification, this pair supports another governance intervention: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions

C:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Map the vendor's model responsibilities separately from the organization's deployment and use controls. The supporting evidence is: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Question 101

An AI project team faces this situation: A transport company celebrates dozens of AI pilots but has no report of unresolved safety concerns. What should the team do first to manage the central issue?

  1. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  2. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  3. Map the vendor's model responsibilities separately from the organization's deployment and use controls.
  4. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

Explanation

Correct Answer

A:

Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Strategic accountability depends on timely risk visibility, not only project-level technical status. The practical acceptance evidence is: review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Incorrect Answers

B:

This would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

C:

This would help address a separate governance task: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

D:

This would help address a separate governance task: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

Question 102

A transport company celebrates dozens of AI pilots but has no report of unresolved safety concerns. Which explanation best connects the risk to the proposed governance action?

  1. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  2. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  3. A generic presentation rarely prepares people to recognize and escalate the failures relevant to their duties.
  4. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Explanation

Correct Answer

B:

Strategic accountability depends on timely risk visibility, not only project-level technical status. That is why the case calls for this action: set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

Incorrect Answers

A:

That explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

C:

That explains why a team might train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

D:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Question 103

A transport company celebrates dozens of AI pilots but has no report of unresolved safety concerns. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  2. Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.
  3. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  4. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Explanation

Correct Answer

C:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Incorrect Answers

A:

Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

B:

Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

D:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Question 104

A team considers this AI use: A transport company celebrates dozens of AI pilots but has no report of unresolved safety concerns. Rather than applying the relevant control, it will report only the number of AI pilots completed as evidence of responsible governance. What can go wrong?

  1. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.
  2. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.
  3. Unknown subcontractors, data rights and change risks can reach production unchecked.
  4. Leadership cannot see whether risks are rising or critical issues remain open.

Explanation

Correct Answer

D:

Leadership cannot see whether risks are rising or critical issues remain open. The shortcut overlooks the central safeguard: set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open.

B:

This consequence follows more directly from another shortcut: Let every involved team assume another team owns the final risk decision. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open.

C:

This consequence follows more directly from another shortcut: Accept a vendor marketing statement as the only evidence of its AI risk controls. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open.

Question 105

A transport company celebrates dozens of AI pilots but has no report of unresolved safety concerns. Which action and evidence pair most directly addresses the situation?

  1. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  2. Action: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. Evidence: Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  3. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Action: Attach explicit review and approval conditions to design, test, release and monitoring milestones. Evidence: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Explanation

Correct Answer

A:

Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions. Together, they implement and test the intended safeguard. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Incorrect Answers

B:

This pair supports another governance intervention: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

C:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

D:

This pair supports another governance intervention: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Question 106

A nonprofit board sees innovation milestones but not the risks accepted for donor-profiling AI. Which step would provide the most targeted control for the problem described?

  1. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.
  2. Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.
  3. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.
  4. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.

Explanation

Correct Answer

B:

For the decision, set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Strategic accountability depends on timely risk visibility, not only project-level technical status. The practical acceptance evidence is: review board or executive reports for risk trends, exceptions, decisions and follow-up actions

Incorrect Answers

A:

This would help address a separate governance task: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

C:

For the decision, this would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions

D:

This would help address a separate governance task: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For this case, however, strategic accountability depends on timely risk visibility, not only project-level technical status. The more direct step is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions.

Question 107

A nonprofit board sees innovation milestones but not the risks accepted for donor-profiling AI. Which explanation best connects the risk to the proposed governance action?

  1. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.
  2. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.
  3. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  4. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Explanation

Correct Answer

C:

For the technical explanation, strategic accountability depends on timely risk visibility, not only project-level technical status. That is why the case calls for this action: set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions

Incorrect Answers

A:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

B:

That explains why a team might apply tiered governance requirements based on impact, organizational capacity and risk tolerance, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

D:

That explains why a team might define triggers, reporting channels, containment authority and post-incident learning for AI harms, but it does not justify the decision required by the case. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Question 108

A nonprofit board sees innovation milestones but not the risks accepted for donor-profiling AI. Which recorded test or observation most directly demonstrates an effective response?

  1. Inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage.
  2. Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  3. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  4. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

D:

For the verification, review board or executive reports for risk trends, exceptions, decisions and follow-up actions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Strategic accountability depends on timely risk visibility, not only project-level technical status

Incorrect Answers

A:

For the verification, inspect a sample of high-risk decisions for named approvers, signoff authority and escalation coverage would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions

B:

Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

C:

Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts would test a different outcome. The directly relevant evidence is to review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Question 109

A nonprofit board sees innovation milestones but not the risks accepted for donor-profiling AI. The team instead proposes: Report only the number of AI pilots completed as evidence of responsible governance. Which failure should the reviewer expect most directly?

  1. Leadership cannot see whether risks are rising or critical issues remain open.
  2. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.
  3. The team may miss legal, societal and practical harms outside its own expertise.
  4. Variable outputs may include failures never observed in the single demonstration.

Explanation

Correct Answer

A:

For the risk comparison, leadership cannot see whether risks are rising or critical issues remain open. The shortcut overlooks the central safeguard: set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Strategic accountability depends on timely risk visibility, not only project-level technical status

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Let every involved team assume another team owns the final risk decision. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open

C:

This consequence follows more directly from another shortcut: Allow only the model builders to review a system that will affect diverse external users. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open.

D:

This consequence follows more directly from another shortcut: Treat one satisfactory model response as proof that all future responses will be equally safe. The shortcut actually described would instead create this risk: Leadership cannot see whether risks are rising or critical issues remain open.

Question 110

Choose the intervention and validation that fit this business problem: A nonprofit board sees innovation milestones but not the risks accepted for donor-profiling AI.

  1. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  2. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.
  3. Action: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. Evidence: Check meeting records for documented challenges, responses, dissent and resulting design changes.
  4. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Explanation

Correct Answer

B:

For the action and verification, set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions. Together, they implement and test the intended safeguard. Strategic accountability depends on timely risk visibility, not only project-level technical status

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions

C:

This pair supports another governance intervention: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

D:

This pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. The supporting evidence is: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Question 111

An AI project team faces this situation: A regional employer requests emotion scoring at interviews without describing who may be affected. What should the team do first to manage the central issue?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  3. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.
  4. Compare relevant error outcomes across affected groups and investigate material disparities before deployment.

Explanation

Correct Answer

C:

Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The practical acceptance evidence is: sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.

Incorrect Answers

A:

This would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

B:

This would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

D:

This would help address a separate governance task: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Question 112

The organization chooses to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level for this situation: A regional employer requests emotion scoring at interviews without describing who may be affected. Which rationale best supports that choice?

  1. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  2. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  3. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.
  4. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Explanation

Correct Answer

D:

Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. That is why the case calls for this action: establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Incorrect Answers

A:

That explains why a team might attach explicit review and approval conditions to design, test, release and monitoring milestones, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

B:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

C:

That explains why a team might map the vendor's model responsibilities separately from the organization's deployment and use controls, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Question 113

A regional employer requests emotion scoring at interviews without describing who may be affected. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  2. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  3. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  4. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

A:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Incorrect Answers

B:

The alternative—check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

C:

The alternative—simulate an unsafe-output incident and verify escalation, containment and corrective documentation—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

D:

The alternative—run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Question 114

The organization has this problem: A regional employer requests emotion scoring at interviews without describing who may be affected. It takes a shortcut: Allow teams to build a model first and identify its intended users only at launch. What is the principal risk?

  1. Changes can bypass controls even when new risks appear in training data or deployment.
  2. A harmful or unsuitable use case may become expensive to correct after development.
  3. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  4. Leadership cannot see whether risks are rising or critical issues remain open.

Explanation

Correct Answer

B:

A harmful or unsuitable use case may become expensive to correct after development. The shortcut overlooks the central safeguard: establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development.

C:

This consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development.

D:

This consequence follows more directly from another shortcut: Report only the number of AI pilots completed as evidence of responsible governance. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development.

Question 115

Choose the intervention and validation that fit this business problem: A regional employer requests emotion scoring at interviews without describing who may be affected.

  1. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  2. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  3. Action: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Evidence: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  4. Action: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

C:

Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented. Together, they implement and test the intended safeguard. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Incorrect Answers

A:

The alternative—require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

B:

This pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. The supporting evidence is: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.

D:

The alternative—assess confidentiality, data minimization and adversarial misuse together in the AI risk register. Evidence: Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Question 116

A school district proposes predictive student interventions without documenting the decisions supported. Which governance action most directly addresses the need in this situation?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  3. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.
  4. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Explanation

Correct Answer

D:

For the decision, establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The practical acceptance evidence is: sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level

B:

For the decision, this would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level

C:

This would help address a separate governance task: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For this case, however, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. The more direct step is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Question 117

A school district proposes predictive student interventions without documenting the decisions supported. Why is the recommended governance approach appropriate?

  1. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  2. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  3. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  4. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Explanation

Correct Answer

A:

For the technical explanation, early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed. That is why the case calls for this action: establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level

Incorrect Answers

B:

For the technical explanation, that explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed

C:

That explains why a team might define triggers, reporting channels, containment authority and post-incident learning for AI harms, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

D:

That explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.

Question 118

A school district proposes predictive student interventions without documenting the decisions supported. Which recorded test or observation most directly demonstrates an effective response?

  1. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  2. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  3. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  4. Check meeting records for documented challenges, responses, dissent and resulting design changes.

Explanation

Correct Answer

B:

For the verification, sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed

Incorrect Answers

A:

The alternative—reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

C:

The alternative—trace a sample release through design review, test evidence, approval and post-launch owner assignments—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

D:

The alternative—check meeting records for documented challenges, responses, dissent and resulting design changes—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

Question 119

A team considers this AI use: A school district proposes predictive student interventions without documenting the decisions supported. Rather than applying the relevant control, it will allow teams to build a model first and identify its intended users only at launch. What can go wrong?

  1. Changes can bypass controls even when new risks appear in training data or deployment.
  2. Ongoing harm may continue because nobody has authority or a clear reporting route.
  3. A harmful or unsuitable use case may become expensive to correct after development.
  4. Leadership cannot see whether risks are rising or critical issues remain open.

Explanation

Correct Answer

C:

For the risk comparison, a harmful or unsuitable use case may become expensive to correct after development. The shortcut overlooks the central safeguard: establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development

B:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Report only the number of AI pilots completed as evidence of responsible governance. The shortcut actually described would instead create this risk: A harmful or unsuitable use case may become expensive to correct after development

Question 120

Choose the intervention and validation that fit this business problem: A school district proposes predictive student interventions without documenting the decisions supported.

  1. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  3. Action: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. Evidence: Use role-based exercises and verify that participants can identify triggers for escalation and human review.
  4. Action: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Evidence: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.

Explanation

Correct Answer

D:

For the action and verification, establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented. Together, they implement and test the intended safeguard. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. The supporting evidence is: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented

B:

The alternative—record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts—is not the most direct response to this case. The matching action is to establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.

C:

This pair supports another governance intervention: Train developers, approvers, operators and frontline users on the distinct AI risks and decisions they control. For the actual case, the required action is: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. The supporting evidence is: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.

Question 121

A fintech team expands its fraud model to new products after a single early compliance review. Which step would provide the most targeted control for the problem described?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.
  3. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  4. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Explanation

Correct Answer

A:

Attach explicit review and approval conditions to design, test, release and monitoring milestones. One-time policy acceptance is insufficient when data, model behavior and operating conditions change. The practical acceptance evidence is: trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Incorrect Answers

B:

This would help address a separate governance task: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

C:

This would help address a separate governance task: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

D:

This would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

Question 122

The organization chooses to attach explicit review and approval conditions to design, test, release and monitoring milestones for this situation: A fintech team expands its fraud model to new products after a single early compliance review. Which rationale best supports that choice?

  1. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  2. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  3. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  4. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Explanation

Correct Answer

B:

One-time policy acceptance is insufficient when data, model behavior and operating conditions change. That is why the case calls for this action: attach explicit review and approval conditions to design, test, release and monitoring milestones.

Incorrect Answers

A:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

C:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

D:

That explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

Question 123

A fintech team expands its fraud model to new products after a single early compliance review. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  2. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  3. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  4. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

C:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments. This verifies the necessary safeguard rather than merely showing that an AI project was completed. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

Incorrect Answers

A:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

B:

Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

D:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Question 124

A fintech team expands its fraud model to new products after a single early compliance review. The team instead proposes: Treat an initial ethics memo as permanent permission for every later model version. Which failure should the reviewer expect most directly?

  1. Ongoing harm may continue because nobody has authority or a clear reporting route.
  2. The organization cannot substantiate who approved a change or what evidence existed.
  3. A well-performing average can conceal severe errors concentrated in one affected group.
  4. Changes can bypass controls even when new risks appear in training data or deployment.

Explanation

Correct Answer

D:

Changes can bypass controls even when new risks appear in training data or deployment. The shortcut overlooks the central safeguard: attach explicit review and approval conditions to design, test, release and monitoring milestones. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment.

B:

This consequence follows more directly from another shortcut: Keep all AI risk approvals in informal chats with no linked version history. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment.

C:

This consequence follows more directly from another shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment.

Question 125

A fintech team expands its fraud model to new products after a single early compliance review. Which action and evidence pair most directly addresses the situation?

  1. Action: Attach explicit review and approval conditions to design, test, release and monitoring milestones. Evidence: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  2. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  3. Action: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Evidence: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  4. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Explanation

Correct Answer

A:

Attach explicit review and approval conditions to design, test, release and monitoring milestones. Trace a sample release through design review, test evidence, approval and post-launch owner assignments. Together, they implement and test the intended safeguard. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

Incorrect Answers

B:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

C:

This pair supports another governance intervention: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

D:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Question 126

An AI project team faces this situation: A workforce platform adds agentic tool actions months after its initial low-risk approval. What should the team do first to manage the central issue?

  1. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  2. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  3. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.
  4. Convene legal, security, domain experts and affected-user representatives before accepting model design decisions.

Explanation

Correct Answer

B:

For the decision, attach explicit review and approval conditions to design, test, release and monitoring milestones. One-time policy acceptance is insufficient when data, model behavior and operating conditions change. The practical acceptance evidence is: trace a sample release through design review, test evidence, approval and post-launch owner assignments

Incorrect Answers

A:

This would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

C:

This would help address a separate governance task: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

D:

This would help address a separate governance task: Convene legal, security, domain experts and affected-user representatives before accepting model design decisions. For this case, however, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. The more direct step is: Attach explicit review and approval conditions to design, test, release and monitoring milestones.

Question 127

A workforce platform adds agentic tool actions months after its initial low-risk approval. Which explanation best connects the risk to the proposed governance action?

  1. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  2. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  3. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  4. Aggregate accuracy can hide systematic harms to subpopulations even without explicit use of a protected attribute.

Explanation

Correct Answer

C:

For the technical explanation, one-time policy acceptance is insufficient when data, model behavior and operating conditions change. That is why the case calls for this action: attach explicit review and approval conditions to design, test, release and monitoring milestones

Incorrect Answers

A:

That explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

B:

For the technical explanation, that explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change

D:

That explains why a team might compare relevant error outcomes across affected groups and investigate material disparities before deployment, but it does not justify the decision required by the case. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.

Question 128

An AI governance reviewer examines this situation: A workforce platform adds agentic tool actions months after its initial low-risk approval. Which acceptance check is most decisive?

  1. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  2. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  3. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  4. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Explanation

Correct Answer

D:

For the verification, trace a sample release through design review, test evidence, approval and post-launch owner assignments. This verifies the necessary safeguard rather than merely showing that an AI project was completed. One-time policy acceptance is insufficient when data, model behavior and operating conditions change

Incorrect Answers

A:

Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

B:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

C:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Question 129

For this case—A workforce platform adds agentic tool actions months after its initial low-risk approval—staff choose the shortcut: Treat an initial ethics memo as permanent permission for every later model version. Which consequence best explains why this is inadequate?

  1. Changes can bypass controls even when new risks appear in training data or deployment.
  2. Unknown subcontractors, data rights and change risks can reach production unchecked.
  3. The organization cannot substantiate who approved a change or what evidence existed.
  4. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

Explanation

Correct Answer

A:

For the risk comparison, changes can bypass controls even when new risks appear in training data or deployment. The shortcut overlooks the central safeguard: attach explicit review and approval conditions to design, test, release and monitoring milestones. One-time policy acceptance is insufficient when data, model behavior and operating conditions change

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Accept a vendor marketing statement as the only evidence of its AI risk controls. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Keep all AI risk approvals in informal chats with no linked version history. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment

D:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: Changes can bypass controls even when new risks appear in training data or deployment.

Question 130

A workforce platform adds agentic tool actions months after its initial low-risk approval. Which action and evidence pair most directly addresses the situation?

  1. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  2. Action: Attach explicit review and approval conditions to design, test, release and monitoring milestones. Evidence: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  3. Action: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Evidence: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  4. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

B:

For the action and verification, attach explicit review and approval conditions to design, test, release and monitoring milestones. Trace a sample release through design review, test evidence, approval and post-launch owner assignments. Together, they implement and test the intended safeguard. One-time policy acceptance is insufficient when data, model behavior and operating conditions change

Incorrect Answers

A:

This pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

C:

For the action and verification, this pair supports another governance intervention: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments

D:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Attach explicit review and approval conditions to design, test, release and monitoring milestones. The supporting evidence is: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.

Question 131

Consider the following proposed AI use: A news publisher drafts an AI principles poster without revisiting article licensing policies. Which response best addresses the governance gap?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  3. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  4. Map the vendor's model responsibilities separately from the organization's deployment and use controls.

Explanation

Correct Answer

C:

Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies. The practical acceptance evidence is: inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Incorrect Answers

A:

The alternative—attach explicit review and approval conditions to design, test, release and monitoring milestones—is not the most direct response to this case. The matching action is to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

B:

This would help address a separate governance task: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For this case, however, existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies. The more direct step is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

D:

The alternative—map the vendor's model responsibilities separately from the organization's deployment and use controls—is not the most direct response to this case. The matching action is to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

Question 132

The organization chooses to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes for this situation: A news publisher drafts an AI principles poster without revisiting article licensing policies. Which rationale best supports that choice?

  1. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  2. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  3. Strategic accountability depends on timely risk visibility, not only project-level technical status.
  4. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Explanation

Correct Answer

D:

Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies. That is why the case calls for this action: update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

Incorrect Answers

A:

That explains why a team might define triggers, reporting channels, containment authority and post-incident learning for AI harms, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

B:

That explains why a team might attach explicit review and approval conditions to design, test, release and monitoring milestones, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

C:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Question 133

An AI governance reviewer examines this situation: A news publisher drafts an AI principles poster without revisiting article licensing policies. Which acceptance check is most decisive?

  1. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  2. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  3. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  4. Review group-level false-positive and false-negative rates with population context and document mitigation decisions.

Explanation

Correct Answer

A:

Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Incorrect Answers

B:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

C:

Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

D:

Review group-level false-positive and false-negative rates with population context and document mitigation decisions would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Question 134

For this case—A news publisher drafts an AI principles poster without revisiting article licensing policies—staff choose the shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. Which consequence best explains why this is inadequate?

  1. Changes can bypass controls even when new risks appear in training data or deployment.
  2. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  3. Ongoing harm may continue because nobody has authority or a clear reporting route.
  4. Resources are wasted on low-risk work while significant risks may still go unidentified.

Explanation

Correct Answer

B:

Employees receive inconsistent instructions for data ingestion, retention and disclosure. The shortcut overlooks the central safeguard: update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure.

C:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure.

D:

This consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure.

Question 135

A news publisher drafts an AI principles poster without revisiting article licensing policies. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  2. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  3. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  4. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.

Explanation

Correct Answer

C:

Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths. Together, they implement and test the intended safeguard. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Incorrect Answers

A:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

B:

This pair supports another governance intervention: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

D:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Question 136

A manufacturer adds AI assistants while its security procedures omit prompts and model connectors. Which governance action most directly addresses the need in this situation?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level.
  3. Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run.
  4. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

Explanation

Correct Answer

D:

For the decision, update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies. The practical acceptance evidence is: inspect amended policies for concrete AI examples, designated owners and consistent escalation paths

Incorrect Answers

A:

For the decision, the alternative—attach explicit review and approval conditions to design, test, release and monitoring milestones—is not the most direct response to this case. The matching action is to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes

B:

The alternative—establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level—is not the most direct response to this case. The matching action is to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

C:

The alternative—set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run—is not the most direct response to this case. The matching action is to update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.

Question 137

A manufacturer adds AI assistants while its security procedures omit prompts and model connectors. Which explanation best connects the risk to the proposed governance action?

  1. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.
  2. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  3. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  4. The supplier may build the model, but the deploying organization still controls its purpose, user access and local monitoring.

Explanation

Correct Answer

A:

For the technical explanation, existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies. That is why the case calls for this action: update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes

Incorrect Answers

B:

That explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

C:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

D:

That explains why a team might map the vendor's model responsibilities separately from the organization's deployment and use controls, but it does not justify the decision required by the case. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.

Question 138

An AI governance reviewer examines this situation: A manufacturer adds AI assistants while its security procedures omit prompts and model connectors. Which acceptance check is most decisive?

  1. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  3. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  4. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

B:

For the verification, inspect amended policies for concrete AI examples, designated owners and consistent escalation paths. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies

Incorrect Answers

A:

For the verification, simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths

C:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

D:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Question 139

The organization has this problem: A manufacturer adds AI assistants while its security procedures omit prompts and model connectors. It takes a shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. What is the principal risk?

  1. Ongoing harm may continue because nobody has authority or a clear reporting route.
  2. A harmful or unsuitable use case may become expensive to correct after development.
  3. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  4. A well-performing average can conceal severe errors concentrated in one affected group.

Explanation

Correct Answer

C:

For the risk comparison, employees receive inconsistent instructions for data ingestion, retention and disclosure. The shortcut overlooks the central safeguard: update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure

B:

This consequence follows more directly from another shortcut: Allow teams to build a model first and identify its intended users only at launch. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure.

D:

This consequence follows more directly from another shortcut: Approve a screening model solely because its overall accuracy exceeds the headline target. The shortcut actually described would instead create this risk: Employees receive inconsistent instructions for data ingestion, retention and disclosure.

Question 140

The governance team needs both a decision and a way to verify it for this case: A manufacturer adds AI assistants while its security procedures omit prompts and model connectors. Which pair is best?

  1. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  3. Action: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. Evidence: Review group-level false-positive and false-negative rates with population context and document mitigation decisions.
  4. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Explanation

Correct Answer

D:

For the action and verification, update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths. Together, they implement and test the intended safeguard. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies

Incorrect Answers

A:

This pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

B:

For the action and verification, this pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths

C:

This pair supports another governance intervention: Compare relevant error outcomes across affected groups and investigate material disparities before deployment. For the actual case, the required action is: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. The supporting evidence is: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.

Question 141

An AI project team faces this situation: A charity selects a fundraising prediction service without asking about training-data sources. What should the team do first to manage the central issue?

  1. Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.
  2. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.
  3. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  4. Apply tiered governance requirements based on impact, organizational capacity and risk tolerance.

Explanation

Correct Answer

A:

Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Governance of a purchased model includes upstream dependencies and limits of contractual assurances. The practical acceptance evidence is: check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Incorrect Answers

B:

This would help address a separate governance task: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

C:

This would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

D:

This would help address a separate governance task: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

Question 142

A team reviewing the following situation asks for the most relevant governance principle: A charity selects a fundraising prediction service without asking about training-data sources. Which principle best explains the required control?

  1. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  2. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  3. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  4. Strategic accountability depends on timely risk visibility, not only project-level technical status.

Explanation

Correct Answer

B:

Governance of a purchased model includes upstream dependencies and limits of contractual assurances. That is why the case calls for this action: require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

Incorrect Answers

A:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

C:

That explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

D:

That explains why a team might set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Question 143

The organization has implemented a control for this case: A charity selects a fundraising prediction service without asking about training-data sources. Which result would best validate it?

  1. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  3. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  4. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

C:

Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Incorrect Answers

A:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

B:

Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

D:

Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Question 144

A charity selects a fundraising prediction service without asking about training-data sources. The team instead proposes: Accept a vendor marketing statement as the only evidence of its AI risk controls. Which failure should the reviewer expect most directly?

  1. Ongoing harm may continue because nobody has authority or a clear reporting route.
  2. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  3. Unauthorized inference-time disclosure or excessive collection can persist despite encryption at rest.
  4. Unknown subcontractors, data rights and change risks can reach production unchecked.

Explanation

Correct Answer

D:

Unknown subcontractors, data rights and change risks can reach production unchecked. The shortcut overlooks the central safeguard: require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

B:

This consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

C:

This consequence follows more directly from another shortcut: Assume encrypted storage alone makes all model inputs and outputs privacy-safe. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

Question 145

The governance team needs both a decision and a way to verify it for this case: A charity selects a fundraising prediction service without asking about training-data sources. Which pair is best?

  1. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  2. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  3. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  4. Action: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. Evidence: Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

A:

Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance. Together, they implement and test the intended safeguard. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Incorrect Answers

B:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

C:

This pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

D:

This pair supports another governance intervention: Set an executive reporting cadence for AI inventory, residual risk and unresolved exceptions. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Question 146

Consider the following proposed AI use: A municipality buys an AI translation platform without investigating external processors or model updates. Which response best addresses the governance gap?

  1. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  2. Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.
  3. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  4. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.

Explanation

Correct Answer

B:

For the decision, require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Governance of a purchased model includes upstream dependencies and limits of contractual assurances. The practical acceptance evidence is: check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance

Incorrect Answers

A:

This would help address a separate governance task: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

C:

For the decision, this would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval

D:

This would help address a separate governance task: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For this case, however, governance of a purchased model includes upstream dependencies and limits of contractual assurances. The more direct step is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.

Question 147

A team reviewing the following situation asks for the most relevant governance principle: A municipality buys an AI translation platform without investigating external processors or model updates. Which principle best explains the required control?

  1. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  2. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  3. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  4. A low-impact recommendation tool and a consequential eligibility decision warrant different levels of evidence and oversight.

Explanation

Correct Answer

C:

For the technical explanation, governance of a purchased model includes upstream dependencies and limits of contractual assurances. That is why the case calls for this action: require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval

Incorrect Answers

A:

For the technical explanation, that explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances

B:

For the technical explanation, that explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances

D:

That explains why a team might apply tiered governance requirements based on impact, organizational capacity and risk tolerance, but it does not justify the decision required by the case. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.

Question 148

The organization has implemented a control for this case: A municipality buys an AI translation platform without investigating external processors or model updates. Which result would best validate it?

  1. Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  2. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  3. Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.
  4. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Explanation

Correct Answer

D:

For the verification, check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance of a purchased model includes upstream dependencies and limits of contractual assurances

Incorrect Answers

A:

For the verification, sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance

B:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

C:

Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance would test a different outcome. The directly relevant evidence is to check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Question 149

A municipality buys an AI translation platform without investigating external processors or model updates. The team instead proposes: Accept a vendor marketing statement as the only evidence of its AI risk controls. Which failure should the reviewer expect most directly?

  1. Unknown subcontractors, data rights and change risks can reach production unchecked.
  2. The organization cannot substantiate who approved a change or what evidence existed.
  3. A harmful or unsuitable use case may become expensive to correct after development.
  4. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.

Explanation

Correct Answer

A:

For the risk comparison, unknown subcontractors, data rights and change risks can reach production unchecked. The shortcut overlooks the central safeguard: require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Governance of a purchased model includes upstream dependencies and limits of contractual assurances

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Keep all AI risk approvals in informal chats with no linked version history. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

C:

This consequence follows more directly from another shortcut: Allow teams to build a model first and identify its intended users only at launch. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

D:

This consequence follows more directly from another shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. The shortcut actually described would instead create this risk: Unknown subcontractors, data rights and change risks can reach production unchecked.

Question 150

The governance team needs both a decision and a way to verify it for this case: A municipality buys an AI translation platform without investigating external processors or model updates. Which pair is best?

  1. Action: Attach explicit review and approval conditions to design, test, release and monitoring milestones. Evidence: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  2. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  3. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  4. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Explanation

Correct Answer

B:

For the action and verification, require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance. Together, they implement and test the intended safeguard. Governance of a purchased model includes upstream dependencies and limits of contractual assurances

Incorrect Answers

A:

This pair supports another governance intervention: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

C:

For the action and verification, this pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance

D:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. The supporting evidence is: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.

Question 151

An AI project team faces this situation: A research institution rotates project leads but has only scattered messages about model approval. What should the team do first to manage the central issue?

  1. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  2. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  3. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.
  4. Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities.

Explanation

Correct Answer

C:

Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reliable audit trails allow accountability and incident investigation after personnel or models change. The practical acceptance evidence is: reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Incorrect Answers

A:

This would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

B:

This would help address a separate governance task: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

D:

This would help address a separate governance task: Publish an AI governance responsibility matrix naming accountable risk owners and approval authorities. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Question 152

The organization chooses to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository for this situation: A research institution rotates project leads but has only scattered messages about model approval. Which rationale best supports that choice?

  1. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  2. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.
  3. Shared participation does not eliminate the need for named accountability for launch, controls and incidents.
  4. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Explanation

Correct Answer

D:

Reliable audit trails allow accountability and incident investigation after personnel or models change. That is why the case calls for this action: record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Incorrect Answers

A:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change.

B:

That explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change.

C:

That explains why a team might publish an AI governance responsibility matrix naming accountable risk owners and approval authorities, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Question 153

A research institution rotates project leads but has only scattered messages about model approval. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  2. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  3. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  4. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

A:

Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Incorrect Answers

B:

The alternative—inspect amended policies for concrete AI examples, designated owners and consistent escalation paths—is not the most direct response to this case. The matching action is to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

C:

The alternative—check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance—is not the most direct response to this case. The matching action is to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

D:

The alternative—review board or executive reports for risk trends, exceptions, decisions and follow-up actions—is not the most direct response to this case. The matching action is to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Question 154

The organization has this problem: A research institution rotates project leads but has only scattered messages about model approval. It takes a shortcut: Keep all AI risk approvals in informal chats with no linked version history. What is the principal risk?

  1. A harmful or unsuitable use case may become expensive to correct after development.
  2. The organization cannot substantiate who approved a change or what evidence existed.
  3. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  4. Material risks can remain unresolved because no person is accountable for accepting or rejecting them.

Explanation

Correct Answer

B:

The organization cannot substantiate who approved a change or what evidence existed. The shortcut overlooks the central safeguard: record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Allow teams to build a model first and identify its intended users only at launch. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed.

C:

This consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed.

D:

This consequence follows more directly from another shortcut: Let every involved team assume another team owns the final risk decision. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed.

Question 155

The governance team needs both a decision and a way to verify it for this case: A research institution rotates project leads but has only scattered messages about model approval. Which pair is best?

  1. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  2. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  3. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  4. Action: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. Evidence: Document whether outputs come from fitted parameters or predefined rules, then map suitable tests to each component.

Explanation

Correct Answer

C:

Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts. Together, they implement and test the intended safeguard. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Incorrect Answers

A:

This pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

B:

This pair supports another governance intervention: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

D:

This pair supports another governance intervention: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Question 156

Consider the following proposed AI use: A public agency loses the test evidence underlying an AI system's last major revision. Which response best addresses the governance gap?

  1. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  2. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  3. Assess confidentiality, data minimization and adversarial misuse together in the AI risk register.
  4. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Explanation

Correct Answer

D:

For the decision, record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reliable audit trails allow accountability and incident investigation after personnel or models change. The practical acceptance evidence is: reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository

B:

For the decision, this would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository

C:

This would help address a separate governance task: Assess confidentiality, data minimization and adversarial misuse together in the AI risk register. For this case, however, reliable audit trails allow accountability and incident investigation after personnel or models change. The more direct step is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Question 157

A public agency loses the test evidence underlying an AI system's last major revision. Why is the recommended governance approach appropriate?

  1. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  2. Existing organizational rules are a starting point but may not cover training data, prompts, model outputs or new dependencies.
  3. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  4. Training and inference can expose personal data or permit misuse even when the model predicts accurately.

Explanation

Correct Answer

A:

For the technical explanation, reliable audit trails allow accountability and incident investigation after personnel or models change. That is why the case calls for this action: record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository

Incorrect Answers

B:

For the technical explanation, that explains why a team might update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change

C:

That explains why a team might define triggers, reporting channels, containment authority and post-incident learning for AI harms, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change.

D:

That explains why a team might assess confidentiality, data minimization and adversarial misuse together in the AI risk register, but it does not justify the decision required by the case. Reliable audit trails allow accountability and incident investigation after personnel or models change.

Question 158

An AI governance reviewer examines this situation: A public agency loses the test evidence underlying an AI system's last major revision. Which acceptance check is most decisive?

  1. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  3. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  4. Test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries.

Explanation

Correct Answer

B:

For the verification, reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reliable audit trails allow accountability and incident investigation after personnel or models change

Incorrect Answers

A:

The alternative—simulate an unsafe-output incident and verify escalation, containment and corrective documentation—is not the most direct response to this case. The matching action is to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

C:

This evidence could test a different control, such as: Attach explicit review and approval conditions to design, test, release and monitoring milestones. It does not establish the safeguard needed here. A decisive check is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

D:

The alternative—test access controls, prompt abuse, output leakage and minimization against approved data-use boundaries—is not the most direct response to this case. The matching action is to record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.

Question 159

A public agency loses the test evidence underlying an AI system's last major revision. The team instead proposes: Keep all AI risk approvals in informal chats with no linked version history. Which failure should the reviewer expect most directly?

  1. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  2. Ongoing harm may continue because nobody has authority or a clear reporting route.
  3. The organization cannot substantiate who approved a change or what evidence existed.
  4. Staff may not know how to respond to unsafe outputs or governance events in their own workflow.

Explanation

Correct Answer

C:

For the risk comparison, the organization cannot substantiate who approved a change or what evidence existed. The shortcut overlooks the central safeguard: record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reliable audit trails allow accountability and incident investigation after personnel or models change

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed

B:

This consequence follows more directly from another shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed.

D:

This consequence follows more directly from another shortcut: Send the same one-page AI definition to every employee and record it as complete competence training. The shortcut actually described would instead create this risk: The organization cannot substantiate who approved a change or what evidence existed.

Question 160

The governance team needs both a decision and a way to verify it for this case: A public agency loses the test evidence underlying an AI system's last major revision. Which pair is best?

  1. Action: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. Evidence: Sample incoming requests to confirm purpose, affected groups, data sources and an initial risk decision are documented.
  2. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  3. Action: Map the vendor's model responsibilities separately from the organization's deployment and use controls. Evidence: Verify signed contracts and runbooks assign model evidence, local oversight, incident contacts and usage restrictions.
  4. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Explanation

Correct Answer

D:

For the action and verification, record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts. Together, they implement and test the intended safeguard. Reliable audit trails allow accountability and incident investigation after personnel or models change

Incorrect Answers

A:

This pair supports another governance intervention: Establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

B:

This pair supports another governance intervention: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

C:

This pair supports another governance intervention: Map the vendor's model responsibilities separately from the organization's deployment and use controls. For the actual case, the required action is: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. The supporting evidence is: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.

Question 161

A social service chatbot repeatedly gives misleading benefits deadlines and no team owns escalation. Which governance action most directly addresses the need in this situation?

  1. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  2. Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository.
  3. Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval.
  4. Classify a text-generating model separately from a numerical risk predictor when preparing evaluations.

Explanation

Correct Answer

A:

Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The practical acceptance evidence is: simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Incorrect Answers

B:

This would help address a separate governance task: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

C:

This would help address a separate governance task: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

D:

This would help address a separate governance task: Classify a text-generating model separately from a numerical risk predictor when preparing evaluations. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

Question 162

A social service chatbot repeatedly gives misleading benefits deadlines and no team owns escalation. Why is the recommended governance approach appropriate?

  1. One-time policy acceptance is insufficient when data, model behavior and operating conditions change.
  2. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  3. Governance of a purchased model includes upstream dependencies and limits of contractual assurances.
  4. Some generative systems vary their responses across runs; safe performance should be measured against stable quality and policy criteria.

Explanation

Correct Answer

B:

Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. That is why the case calls for this action: define triggers, reporting channels, containment authority and post-incident learning for AI harms.

Incorrect Answers

A:

That explains why a team might attach explicit review and approval conditions to design, test, release and monitoring milestones, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

C:

That explains why a team might require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

D:

That explains why a team might set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Question 163

An AI governance reviewer examines this situation: A social service chatbot repeatedly gives misleading benefits deadlines and no team owns escalation. Which acceptance check is most decisive?

  1. Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  2. Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  3. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  4. Review board or executive reports for risk trends, exceptions, decisions and follow-up actions.

Explanation

Correct Answer

C:

Simulate an unsafe-output incident and verify escalation, containment and corrective documentation. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Incorrect Answers

A:

Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

B:

Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

D:

Review board or executive reports for risk trends, exceptions, decisions and follow-up actions would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Question 164

The organization has this problem: A social service chatbot repeatedly gives misleading benefits deadlines and no team owns escalation. It takes a shortcut: Wait for an annual policy review before responding to a suspected unsafe AI decision. What is the principal risk?

  1. Changes can bypass controls even when new risks appear in training data or deployment.
  2. The organization cannot substantiate who approved a change or what evidence existed.
  3. Local misuse, weak access control or poor monitoring can remain unmanaged by either party.
  4. Ongoing harm may continue because nobody has authority or a clear reporting route.

Explanation

Correct Answer

D:

Ongoing harm may continue because nobody has authority or a clear reporting route. The shortcut overlooks the central safeguard: define triggers, reporting channels, containment authority and post-incident learning for AI harms. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat an initial ethics memo as permanent permission for every later model version. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

B:

This consequence follows more directly from another shortcut: Keep all AI risk approvals in informal chats with no linked version history. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

C:

This consequence follows more directly from another shortcut: Assume purchasing an AI service transfers every operational governance duty to the supplier. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

Question 165

The governance team needs both a decision and a way to verify it for this case: A social service chatbot repeatedly gives misleading benefits deadlines and no team owns escalation. Which pair is best?

  1. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  2. Action: Attach explicit review and approval conditions to design, test, release and monitoring milestones. Evidence: Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  3. Action: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. Evidence: Check procurement files for completed due diligence, negotiated controls and unresolved risk acceptance.
  4. Action: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. Evidence: Compare low- and high-impact project dossiers to confirm proportionate controls and justified exceptions.

Explanation

Correct Answer

A:

Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation. Together, they implement and test the intended safeguard. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Incorrect Answers

B:

This pair supports another governance intervention: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

C:

This pair supports another governance intervention: Require vendor evidence on data use, subcontractors, security, model changes and exit arrangements before procurement approval. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

D:

This pair supports another governance intervention: Apply tiered governance requirements based on impact, organizational capacity and risk tolerance. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Question 166

An AI project team faces this situation: A retailer discovers its hiring screener systematically rejects valid candidates during a busy intake cycle. What should the team do first to manage the central issue?

  1. Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes.
  2. Define triggers, reporting channels, containment authority and post-incident learning for AI harms.
  3. Attach explicit review and approval conditions to design, test, release and monitoring milestones.
  4. Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls.

Explanation

Correct Answer

B:

For the decision, define triggers, reporting channels, containment authority and post-incident learning for AI harms. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The practical acceptance evidence is: simulate an unsafe-output incident and verify escalation, containment and corrective documentation

Incorrect Answers

A:

This would help address a separate governance task: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

C:

This would help address a separate governance task: Attach explicit review and approval conditions to design, test, release and monitoring milestones. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

D:

This would help address a separate governance task: Distinguish learned statistical prediction from a fixed rules engine before assigning AI-specific controls. For this case, however, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. The more direct step is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms.

Question 167

A retailer discovers its hiring screener systematically rejects valid candidates during a busy intake cycle. Why is the recommended governance approach appropriate?

  1. Early structured intake makes prohibited, excessive or poorly scoped uses visible before resources are committed.
  2. Reliable audit trails allow accountability and incident investigation after personnel or models change.
  3. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.
  4. Different disciplines reveal failure modes that a purely technical or purely legal review may overlook.

Explanation

Correct Answer

C:

For the technical explanation, ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately. That is why the case calls for this action: define triggers, reporting channels, containment authority and post-incident learning for AI harms

Incorrect Answers

A:

That explains why a team might establish an AI use-case intake record capturing purpose, stakeholders, data, impacts and preliminary risk level, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

B:

That explains why a team might record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

D:

That explains why a team might convene legal, security, domain experts and affected-user representatives before accepting model design decisions, but it does not justify the decision required by the case. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately.

Question 168

The organization has implemented a control for this case: A retailer discovers its hiring screener systematically rejects valid candidates during a busy intake cycle. Which result would best validate it?

  1. Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  2. Trace a sample release through design review, test evidence, approval and post-launch owner assignments.
  3. Review samples of generated text for supported claims and separately test predictor error on held-out cases.
  4. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Explanation

Correct Answer

D:

For the verification, simulate an unsafe-output incident and verify escalation, containment and corrective documentation. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately

Incorrect Answers

A:

Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

B:

Trace a sample release through design review, test evidence, approval and post-launch owner assignments would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

C:

Review samples of generated text for supported claims and separately test predictor error on held-out cases would test a different outcome. The directly relevant evidence is to simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Question 169

A team considers this AI use: A retailer discovers its hiring screener systematically rejects valid candidates during a busy intake cycle. Rather than applying the relevant control, it will wait for an annual policy review before responding to a suspected unsafe AI decision. What can go wrong?

  1. Ongoing harm may continue because nobody has authority or a clear reporting route.
  2. Employees receive inconsistent instructions for data ingestion, retention and disclosure.
  3. A harmful or unsuitable use case may become expensive to correct after development.
  4. Resources are wasted on low-risk work while significant risks may still go unidentified.

Explanation

Correct Answer

A:

For the risk comparison, ongoing harm may continue because nobody has authority or a clear reporting route. The shortcut overlooks the central safeguard: define triggers, reporting channels, containment authority and post-incident learning for AI harms. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Create a separate AI slogan while leaving conflicting data retention and access rules unchanged. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

C:

This consequence follows more directly from another shortcut: Allow teams to build a model first and identify its intended users only at launch. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

D:

This consequence follows more directly from another shortcut: Give an experimental calendar helper the same gate as an automated loan decision while skipping both systems' risk assessment. The shortcut actually described would instead create this risk: Ongoing harm may continue because nobody has authority or a clear reporting route.

Question 170

A retailer discovers its hiring screener systematically rejects valid candidates during a busy intake cycle. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. Evidence: Inspect amended policies for concrete AI examples, designated owners and consistent escalation paths.
  2. Action: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. Evidence: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.
  3. Action: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. Evidence: Reconstruct a sample system's decisions from dated records, assigned owners and linked artifacts.
  4. Action: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. Evidence: Run multiple controlled prompts and record factual accuracy, policy compliance and acceptable variance.

Explanation

Correct Answer

B:

For the action and verification, define triggers, reporting channels, containment authority and post-incident learning for AI harms. Simulate an unsafe-output incident and verify escalation, containment and corrective documentation. Together, they implement and test the intended safeguard. Ordinary IT tickets may not capture biased outcomes, generated unsafe advice or model misuse appropriately

Incorrect Answers

A:

This pair supports another governance intervention: Update privacy, cybersecurity, records and IP policies to address AI-specific workflows and failure modes. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

C:

This pair supports another governance intervention: Record the system's intended purpose, risk assessments, decisions, changes and evidence in a controlled repository. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

D:

This pair supports another governance intervention: Set repeatability tolerances and sample-based tests for a probabilistic AI service instead of requiring identical wording on every run. For the actual case, the required action is: Define triggers, reporting channels, containment authority and post-incident learning for AI harms. The supporting evidence is: Simulate an unsafe-output incident and verify escalation, containment and corrective documentation.

Question 171

A European insurer wants to train a churn model on policyholder claims records collected for servicing. Which step would provide the most targeted control for the problem described?

  1. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  2. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  3. Identify and document the applicable lawful basis before reusing personal information for model development.
  4. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

Explanation

Correct Answer

C:

Identify and document the applicable lawful basis before reusing personal information for model development. Having technical access to records does not, by itself, authorize a new AI training purpose. The practical acceptance evidence is: check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Incorrect Answers

A:

This would help address a separate governance task: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For this case, however, having technical access to records does not, by itself, authorize a new AI training purpose. The more direct step is: Identify and document the applicable lawful basis before reusing personal information for model development.

B:

The alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—is not the most direct response to this case. The matching action is to identify and document the applicable lawful basis before reusing personal information for model development.

D:

This would help address a separate governance task: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For this case, however, having technical access to records does not, by itself, authorize a new AI training purpose. The more direct step is: Identify and document the applicable lawful basis before reusing personal information for model development.

Question 172

The organization chooses to identify and document the applicable lawful basis before reusing personal information for model development for this situation: A European insurer wants to train a churn model on policyholder claims records collected for servicing. Which rationale best supports that choice?

  1. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  2. Collecting more personal data than required increases exposure without automatically improving the model.
  3. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.
  4. Having technical access to records does not, by itself, authorize a new AI training purpose.

Explanation

Correct Answer

D:

Having technical access to records does not, by itself, authorize a new AI training purpose. That is why the case calls for this action: identify and document the applicable lawful basis before reusing personal information for model development.

Incorrect Answers

A:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose.

B:

That explains why a team might remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose.

C:

That explains why a team might check the provider's contract and relevant law for permitted commercial use and responsibility for generated content, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose.

Question 173

An AI governance reviewer examines this situation: A European insurer wants to train a churn model on policyholder claims records collected for servicing. Which acceptance check is most decisive?

  1. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  2. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  4. Review license terms and approval records for representative public-facing generated materials.

Explanation

Correct Answer

A:

Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Having technical access to records does not, by itself, authorize a new AI training purpose.

Incorrect Answers

B:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

C:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

D:

Review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Question 174

A European insurer wants to train a churn model on policyholder claims records collected for servicing. The team instead proposes: Assume a dataset may be trained on because staff already can open the underlying system. Which failure should the reviewer expect most directly?

  1. People may lose legally required recourse or receive no usable explanation of a consequential outcome.
  2. An unauthorized new processing purpose can be hidden behind existing access rights.
  3. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  4. Public outputs can create contractual or infringement exposure even without intentional copying.

Explanation

Correct Answer

B:

An unauthorized new processing purpose can be hidden behind existing access rights. The shortcut overlooks the central safeguard: identify and document the applicable lawful basis before reusing personal information for model development. Having technical access to records does not, by itself, authorize a new AI training purpose.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Tell rejected applicants that model decisions can never be reviewed because the system is automated. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights.

C:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights.

D:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights.

Question 175

A European insurer wants to train a churn model on policyholder claims records collected for servicing. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  2. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  3. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  4. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Explanation

Correct Answer

C:

Identify and document the applicable lawful basis before reusing personal information for model development. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion. Together, they implement and test the intended safeguard. Having technical access to records does not, by itself, authorize a new AI training purpose.

Incorrect Answers

A:

This pair supports another governance intervention: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

B:

This pair supports another governance intervention: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

D:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Question 176

A research hospital plans to reuse appointment histories for a new prediction task without a privacy review. Which step would provide the most targeted control for the problem described?

  1. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  2. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  3. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  4. Identify and document the applicable lawful basis before reusing personal information for model development.

Explanation

Correct Answer

D:

For the decision, identify and document the applicable lawful basis before reusing personal information for model development. Having technical access to records does not, by itself, authorize a new AI training purpose. The practical acceptance evidence is: check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For this case, however, having technical access to records does not, by itself, authorize a new AI training purpose. The more direct step is: Identify and document the applicable lawful basis before reusing personal information for model development

B:

For the decision, the alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—is not the most direct response to this case. The matching action is to identify and document the applicable lawful basis before reusing personal information for model development

C:

This would help address a separate governance task: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For this case, however, having technical access to records does not, by itself, authorize a new AI training purpose. The more direct step is: Identify and document the applicable lawful basis before reusing personal information for model development.

Question 177

The organization chooses to identify and document the applicable lawful basis before reusing personal information for model development for this situation: A research hospital plans to reuse appointment histories for a new prediction task without a privacy review. Which rationale best supports that choice?

  1. Having technical access to records does not, by itself, authorize a new AI training purpose.
  2. Collecting more personal data than required increases exposure without automatically improving the model.
  3. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  4. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Explanation

Correct Answer

A:

For the technical explanation, having technical access to records does not, by itself, authorize a new AI training purpose. That is why the case calls for this action: identify and document the applicable lawful basis before reusing personal information for model development

Incorrect Answers

B:

For the technical explanation, that explains why a team might remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose

C:

For the technical explanation, that explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose

D:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Having technical access to records does not, by itself, authorize a new AI training purpose.

Question 178

A research hospital plans to reuse appointment histories for a new prediction task without a privacy review. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  2. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  4. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

B:

For the verification, check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Having technical access to records does not, by itself, authorize a new AI training purpose

Incorrect Answers

A:

Compare approved feature requirements with actual datasets and test a reduced-data version for utility would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

C:

Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

D:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Question 179

The organization has this problem: A research hospital plans to reuse appointment histories for a new prediction task without a privacy review. It takes a shortcut: Assume a dataset may be trained on because staff already can open the underlying system. What is the principal risk?

  1. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  2. People's data may support decisions they were never reasonably told to expect.
  3. An unauthorized new processing purpose can be hidden behind existing access rights.
  4. The organization may build a use that cannot lawfully operate in its target context.

Explanation

Correct Answer

C:

For the risk comparison, an unauthorized new processing purpose can be hidden behind existing access rights. The shortcut overlooks the central safeguard: identify and document the applicable lawful basis before reusing personal information for model development. Having technical access to records does not, by itself, authorize a new AI training purpose

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights

B:

This consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights.

D:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: An unauthorized new processing purpose can be hidden behind existing access rights.

Question 180

Choose the intervention and validation that fit this business problem: A research hospital plans to reuse appointment histories for a new prediction task without a privacy review.

  1. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.
  4. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Explanation

Correct Answer

D:

For the action and verification, identify and document the applicable lawful basis before reusing personal information for model development. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion. Together, they implement and test the intended safeguard. Having technical access to records does not, by itself, authorize a new AI training purpose

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion

B:

This pair supports another governance intervention: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

C:

This pair supports another governance intervention: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For the actual case, the required action is: Identify and document the applicable lawful basis before reusing personal information for model development. The supporting evidence is: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.

Question 181

Consider the following proposed AI use: A transit operator wants to use travel-card logs collected for ticketing to infer employee productivity. Which response best addresses the governance gap?

  1. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  2. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  3. Identify and document the applicable lawful basis before reusing personal information for model development.
  4. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.

Explanation

Correct Answer

A:

Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards. The practical acceptance evidence is: record the original notice and collection purpose alongside the proposed use and compatibility determination.

Incorrect Answers

B:

The alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

C:

The alternative—identify and document the applicable lawful basis before reusing personal information for model development—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 182

A transit operator wants to use travel-card logs collected for ticketing to infer employee productivity. Why is the recommended governance approach appropriate?

  1. Having technical access to records does not, by itself, authorize a new AI training purpose.
  2. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.
  3. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  4. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Explanation

Correct Answer

B:

Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards. That is why the case calls for this action: test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.

Incorrect Answers

A:

The alternative—having technical access to records does not, by itself, authorize a new AI training purpose—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

C:

The alternative—aI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 183

A transit operator wants to use travel-card logs collected for ticketing to infer employee productivity. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  3. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  4. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Explanation

Correct Answer

C:

Record the original notice and collection purpose alongside the proposed use and compatibility determination. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Incorrect Answers

A:

The alternative—test sample requests for access, correction and review against the governing jurisdictional rules and operational logs—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

B:

The alternative—confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—trace a selected use case from contextual mapping to measured evidence and a documented risk response—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 184

The organization has this problem: A transit operator wants to use travel-card logs collected for ticketing to infer employee productivity. It takes a shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. What is the principal risk?

  1. An unauthorized new processing purpose can be hidden behind existing access rights.
  2. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  3. A role change may leave mandatory obligations or documentation unassigned.
  4. People's data may support decisions they were never reasonably told to expect.

Explanation

Correct Answer

D:

People's data may support decisions they were never reasonably told to expect. The shortcut overlooks the central safeguard: test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect.

B:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect.

C:

This consequence follows more directly from another shortcut: Assume every organization interacting with an AI system is legally only a passive end user. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect.

Question 185

Choose the intervention and validation that fit this business problem: A transit operator wants to use travel-card logs collected for ticketing to infer employee productivity.

  1. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  2. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  3. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  4. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

A:

Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Record the original notice and collection purpose alongside the proposed use and compatibility determination. Together, they implement and test the intended safeguard. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Incorrect Answers

B:

The alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

C:

The alternative—design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 186

A nonprofit plans to use donor contact records collected for receipts to score volunteers for eligibility. Which governance action most directly addresses the need in this situation?

  1. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  2. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  3. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  4. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

Explanation

Correct Answer

B:

For the decision, test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards. The practical acceptance evidence is: record the original notice and collection purpose alongside the proposed use and compatibility determination

Incorrect Answers

A:

For the decision, the alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

C:

The alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 187

A nonprofit plans to use donor contact records collected for receipts to score volunteers for eligibility. Which explanation best connects the risk to the proposed governance action?

  1. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  2. Collecting more personal data than required increases exposure without automatically improving the model.
  3. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.
  4. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Explanation

Correct Answer

C:

For the technical explanation, reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards. That is why the case calls for this action: test whether a proposed new model purpose is compatible with the reason personal information was originally obtained

Incorrect Answers

A:

For the technical explanation, the alternative—aI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

B:

The alternative—collecting more personal data than required increases exposure without automatically improving the model—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 188

A nonprofit plans to use donor contact records collected for receipts to score volunteers for eligibility. Which recorded test or observation most directly demonstrates an effective response?

  1. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  2. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Record the original notice and collection purpose alongside the proposed use and compatibility determination.

Explanation

Correct Answer

D:

For the verification, record the original notice and collection purpose alongside the proposed use and compatibility determination. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

Incorrect Answers

A:

For the verification, the alternative—confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

B:

The alternative—check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

C:

The alternative—map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 189

The organization has this problem: A nonprofit plans to use donor contact records collected for receipts to score volunteers for eligibility. It takes a shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. What is the principal risk?

  1. People's data may support decisions they were never reasonably told to expect.
  2. An unauthorized new processing purpose can be hidden behind existing access rights.
  3. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  4. A dangerous design defect or inadequate warning may go unaddressed.

Explanation

Correct Answer

A:

For the risk comparison, people's data may support decisions they were never reasonably told to expect. The shortcut overlooks the central safeguard: test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect

C:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect.

D:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: People's data may support decisions they were never reasonably told to expect.

Question 190

The governance team needs both a decision and a way to verify it for this case: A nonprofit plans to use donor contact records collected for receipts to score volunteers for eligibility. Which pair is best?

  1. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  4. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.

Explanation

Correct Answer

B:

For the action and verification, test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Record the original notice and collection purpose alongside the proposed use and compatibility determination. Together, they implement and test the intended safeguard. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

Incorrect Answers

A:

For the action and verification, the alternative—design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards

C:

The alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

D:

The alternative—check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials—would not resolve the stated issue. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Question 191

A hotel is developing a service-demand forecast using bookings but considers ingesting passport numbers. Which step would provide the most targeted control for the problem described?

  1. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  2. Identify and document the applicable lawful basis before reusing personal information for model development.
  3. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  4. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Explanation

Correct Answer

C:

Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Collecting more personal data than required increases exposure without automatically improving the model. The practical acceptance evidence is: compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Incorrect Answers

A:

This would help address a separate governance task: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

B:

This would help address a separate governance task: Identify and document the applicable lawful basis before reusing personal information for model development. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

D:

This would help address a separate governance task: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

Question 192

The organization chooses to remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient for this situation: A hotel is developing a service-demand forecast using bookings but considers ingesting passport numbers. Which rationale best supports that choice?

  1. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  2. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.
  3. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  4. Collecting more personal data than required increases exposure without automatically improving the model.

Explanation

Correct Answer

D:

Collecting more personal data than required increases exposure without automatically improving the model. That is why the case calls for this action: remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

Incorrect Answers

A:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model.

B:

That explains why a team might test whether a proposed new model purpose is compatible with the reason personal information was originally obtained, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model.

C:

That explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model.

Question 193

A hotel is developing a service-demand forecast using bookings but considers ingesting passport numbers. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  2. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  4. Check the legal classification and an obligation-by-obligation evidence matrix before release.

Explanation

Correct Answer

A:

Compare approved feature requirements with actual datasets and test a reduced-data version for utility. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Collecting more personal data than required increases exposure without automatically improving the model.

Incorrect Answers

B:

Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility.

C:

Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility.

D:

Check the legal classification and an obligation-by-obligation evidence matrix before release would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Question 194

The organization has this problem: A hotel is developing a service-demand forecast using bookings but considers ingesting passport numbers. It takes a shortcut: Feed the entire customer database into the model because storage and compute are available. What is the principal risk?

  1. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  2. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  3. An unauthorized new processing purpose can be hidden behind existing access rights.
  4. The organization may be unable to show compliance or respond effectively to a formal inquiry.

Explanation

Correct Answer

B:

Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit. The shortcut overlooks the central safeguard: remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Collecting more personal data than required increases exposure without automatically improving the model.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

C:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

D:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

Question 195

A hotel is developing a service-demand forecast using bookings but considers ingesting passport numbers. Which action and evidence pair most directly addresses the situation?

  1. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  2. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  4. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Explanation

Correct Answer

C:

Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Compare approved feature requirements with actual datasets and test a reduced-data version for utility. Together, they implement and test the intended safeguard. Collecting more personal data than required increases exposure without automatically improving the model.

Incorrect Answers

A:

This pair supports another governance intervention: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

B:

This pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

D:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Question 196

An education provider wants a help-desk classifier and proposes adding pupils' home addresses. Which governance action most directly addresses the need in this situation?

  1. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  2. Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.
  3. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  4. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

Explanation

Correct Answer

D:

For the decision, remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Collecting more personal data than required increases exposure without automatically improving the model. The practical acceptance evidence is: compare approved feature requirements with actual datasets and test a reduced-data version for utility

Incorrect Answers

A:

This would help address a separate governance task: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

B:

This would help address a separate governance task: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

C:

This would help address a separate governance task: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For this case, however, collecting more personal data than required increases exposure without automatically improving the model. The more direct step is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

Question 197

A team reviewing the following situation asks for the most relevant governance principle: An education provider wants a help-desk classifier and proposes adding pupils' home addresses. Which principle best explains the required control?

  1. Collecting more personal data than required increases exposure without automatically improving the model.
  2. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  3. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  4. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

Explanation

Correct Answer

A:

For the technical explanation, collecting more personal data than required increases exposure without automatically improving the model. That is why the case calls for this action: remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient

Incorrect Answers

B:

That explains why a team might design a process to identify relevant data subject requests and assess applicable automated-decision safeguards, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model.

C:

For the technical explanation, that explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model

D:

That explains why a team might conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development, but it does not justify the decision required by the case. Collecting more personal data than required increases exposure without automatically improving the model.

Question 198

The organization has implemented a control for this case: An education provider wants a help-desk classifier and proposes adding pupils' home addresses. Which result would best validate it?

  1. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  3. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  4. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Explanation

Correct Answer

B:

For the verification, compare approved feature requirements with actual datasets and test a reduced-data version for utility. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Collecting more personal data than required increases exposure without automatically improving the model

Incorrect Answers

A:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility.

C:

For the verification, confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility

D:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Question 199

The organization has this problem: An education provider wants a help-desk classifier and proposes adding pupils' home addresses. It takes a shortcut: Feed the entire customer database into the model because storage and compute are available. What is the principal risk?

  1. People may lose legally required recourse or receive no usable explanation of a consequential outcome.
  2. An unauthorized new processing purpose can be hidden behind existing access rights.
  3. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  4. Public outputs can create contractual or infringement exposure even without intentional copying.

Explanation

Correct Answer

C:

For the risk comparison, unneeded sensitive fields can leak or influence outcomes while adding no justified benefit. The shortcut overlooks the central safeguard: remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Collecting more personal data than required increases exposure without automatically improving the model

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Tell rejected applicants that model decisions can never be reviewed because the system is automated. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit

D:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

Question 200

The governance team needs both a decision and a way to verify it for this case: An education provider wants a help-desk classifier and proposes adding pupils' home addresses. Which pair is best?

  1. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  2. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  3. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Explanation

Correct Answer

D:

For the action and verification, remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Compare approved feature requirements with actual datasets and test a reduced-data version for utility. Together, they implement and test the intended safeguard. Collecting more personal data than required increases exposure without automatically improving the model

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility

B:

This pair supports another governance intervention: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

C:

This pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. The supporting evidence is: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Question 201

Consider the following proposed AI use: A stadium operator proposes facial matching for entry without a dedicated privacy assessment. Which response best addresses the governance gap?

  1. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  2. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  3. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  4. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Explanation

Correct Answer

A:

Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. The practical acceptance evidence is: confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Incorrect Answers

B:

The alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—would not resolve the stated issue. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

C:

The alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—would not resolve the stated issue. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

D:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—would not resolve the stated issue. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Question 202

A stadium operator proposes facial matching for entry without a dedicated privacy assessment. Which explanation best connects the risk to the proposed governance action?

  1. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  2. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  3. Collecting more personal data than required increases exposure without automatically improving the model.
  4. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Explanation

Correct Answer

B:

Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. That is why the case calls for this action: escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.

Incorrect Answers

A:

That explains why a team might design a process to identify relevant data subject requests and assess applicable automated-decision safeguards, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

C:

That explains why a team might remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

D:

That explains why a team might use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Question 203

The organization has implemented a control for this case: A stadium operator proposes facial matching for entry without a dedicated privacy assessment. Which result would best validate it?

  1. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  2. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  3. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  4. Review provider technical documentation and downstream integration requirements under the relevant regime.

Explanation

Correct Answer

C:

Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Incorrect Answers

A:

Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

B:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

D:

Review provider technical documentation and downstream integration requirements under the relevant regime would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Question 204

A stadium operator proposes facial matching for entry without a dedicated privacy assessment. The team instead proposes: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. Which failure should the reviewer expect most directly?

  1. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  2. An unauthorized new processing purpose can be hidden behind existing access rights.
  3. The organization may build a use that cannot lawfully operate in its target context.
  4. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

Explanation

Correct Answer

D:

Identifiable or sensitive characteristics may still be exposed or processed unlawfully. The shortcut overlooks the central safeguard: escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

B:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

C:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

Question 205

A stadium operator proposes facial matching for entry without a dedicated privacy assessment. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  2. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  4. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

A:

Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions. Together, they implement and test the intended safeguard. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Incorrect Answers

B:

This pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

C:

This pair supports another governance intervention: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

D:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Question 206

A health startup stores voice biometrics for account access and assumes pseudonyms remove privacy duties. Which governance action most directly addresses the need in this situation?

  1. Identify and document the applicable lawful basis before reusing personal information for model development.
  2. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  3. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  4. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

Explanation

Correct Answer

B:

For the decision, escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. The practical acceptance evidence is: confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions

Incorrect Answers

A:

This would help address a separate governance task: Identify and document the applicable lawful basis before reusing personal information for model development. For this case, however, sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. The more direct step is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.

C:

For the decision, the alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—would not resolve the stated issue. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses

D:

This would help address a separate governance task: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For this case, however, sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. The more direct step is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.

Question 207

A team reviewing the following situation asks for the most relevant governance principle: A health startup stores voice biometrics for account access and assumes pseudonyms remove privacy duties. Which principle best explains the required control?

  1. Having technical access to records does not, by itself, authorize a new AI training purpose.
  2. Collecting more personal data than required increases exposure without automatically improving the model.
  3. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  4. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Explanation

Correct Answer

C:

For the technical explanation, sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses. That is why the case calls for this action: escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review

Incorrect Answers

A:

That explains why a team might identify and document the applicable lawful basis before reusing personal information for model development, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

B:

For the technical explanation, that explains why a team might remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses

D:

That explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Question 208

A health startup stores voice biometrics for account access and assumes pseudonyms remove privacy duties. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Explanation

Correct Answer

D:

For the verification, confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses

Incorrect Answers

A:

For the verification, test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions

B:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

C:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Question 209

A team considers this AI use: A health startup stores voice biometrics for account access and assumes pseudonyms remove privacy duties. Rather than applying the relevant control, it will treat face embeddings as ordinary anonymous telemetry simply because names are removed. What can go wrong?

  1. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  2. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  3. People may lose legally required recourse or receive no usable explanation of a consequential outcome.
  4. The organization can overlook use context and fail to act on significant measured risks.

Explanation

Correct Answer

A:

For the risk comparison, identifiable or sensitive characteristics may still be exposed or processed unlawfully. The shortcut overlooks the central safeguard: escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully

C:

This consequence follows more directly from another shortcut: Tell rejected applicants that model decisions can never be reviewed because the system is automated. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

D:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

Question 210

Choose the intervention and validation that fit this business problem: A health startup stores voice biometrics for account access and assumes pseudonyms remove privacy duties.

  1. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  2. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  3. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  4. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.

Explanation

Correct Answer

B:

For the action and verification, escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions. Together, they implement and test the intended safeguard. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions

C:

This pair supports another governance intervention: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

D:

This pair supports another governance intervention: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For the actual case, the required action is: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. The supporting evidence is: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Question 211

An AI project team faces this situation: A bank deploys automated account flags and lacks a process for customers to challenge incorrect personal data. What should the team do first to manage the central issue?

  1. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  2. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.
  3. Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.
  4. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

Explanation

Correct Answer

C:

Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. The practical acceptance evidence is: test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Incorrect Answers

A:

The alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—is not the most direct response to this case. The matching action is to design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

B:

The alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—is not the most direct response to this case. The matching action is to design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

D:

The alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review—is not the most direct response to this case. The matching action is to design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Question 212

A team reviewing the following situation asks for the most relevant governance principle: A bank deploys automated account flags and lacks a process for customers to challenge incorrect personal data. Which principle best explains the required control?

  1. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  2. Having technical access to records does not, by itself, authorize a new AI training purpose.
  3. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  4. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Explanation

Correct Answer

D:

AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. That is why the case calls for this action: design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Incorrect Answers

A:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

B:

That explains why a team might identify and document the applicable lawful basis before reusing personal information for model development, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

C:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Question 213

A bank deploys automated account flags and lacks a process for customers to challenge incorrect personal data. Which recorded test or observation most directly demonstrates an effective response?

  1. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  2. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  4. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Explanation

Correct Answer

A:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs. This verifies the necessary safeguard rather than merely showing that an AI project was completed. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Incorrect Answers

B:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

C:

Compare approved feature requirements with actual datasets and test a reduced-data version for utility would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

D:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Question 214

For this case—A bank deploys automated account flags and lacks a process for customers to challenge incorrect personal data—staff choose the shortcut: Tell rejected applicants that model decisions can never be reviewed because the system is automated. Which consequence best explains why this is inadequate?

  1. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.
  2. People may lose legally required recourse or receive no usable explanation of a consequential outcome.
  3. People's data may support decisions they were never reasonably told to expect.
  4. The organization may be unable to show compliance or respond effectively to a formal inquiry.

Explanation

Correct Answer

B:

People may lose legally required recourse or receive no usable explanation of a consequential outcome. The shortcut overlooks the central safeguard: design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome.

C:

This consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome.

D:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome.

Question 215

The governance team needs both a decision and a way to verify it for this case: A bank deploys automated account flags and lacks a process for customers to challenge incorrect personal data. Which pair is best?

  1. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  2. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  3. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  4. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

C:

Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs. Together, they implement and test the intended safeguard. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Incorrect Answers

A:

This pair supports another governance intervention: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

B:

This pair supports another governance intervention: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

D:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Question 216

An AI project team faces this situation: A recruitment platform provides rejection scores without checking obligations for human review and candidate requests. What should the team do first to manage the central issue?

  1. Identify and document the applicable lawful basis before reusing personal information for model development.
  2. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  3. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.
  4. Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Explanation

Correct Answer

D:

For the decision, design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. The practical acceptance evidence is: test sample requests for access, correction and review against the governing jurisdictional rules and operational logs

Incorrect Answers

A:

This would help address a separate governance task: Identify and document the applicable lawful basis before reusing personal information for model development. For this case, however, aI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. The more direct step is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

B:

For the decision, the alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—is not the most direct response to this case. The matching action is to design a process to identify relevant data subject requests and assess applicable automated-decision safeguards

C:

This would help address a separate governance task: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For this case, however, aI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. The more direct step is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Question 217

A recruitment platform provides rejection scores without checking obligations for human review and candidate requests. Why is the recommended governance approach appropriate?

  1. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  2. Having technical access to records does not, by itself, authorize a new AI training purpose.
  3. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.
  4. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Explanation

Correct Answer

A:

For the technical explanation, aI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable. That is why the case calls for this action: design a process to identify relevant data subject requests and assess applicable automated-decision safeguards

Incorrect Answers

B:

For the technical explanation, that explains why a team might identify and document the applicable lawful basis before reusing personal information for model development, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable

C:

That explains why a team might test whether a proposed new model purpose is compatible with the reason personal information was originally obtained, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

D:

That explains why a team might review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use, but it does not justify the decision required by the case. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Question 218

A recruitment platform provides rejection scores without checking obligations for human review and candidate requests. Which recorded test or observation most directly demonstrates an effective response?

  1. Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.
  2. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  3. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  4. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

B:

For the verification, test sample requests for access, correction and review against the governing jurisdictional rules and operational logs. This verifies the necessary safeguard rather than merely showing that an AI project was completed. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable

Incorrect Answers

A:

Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

C:

For the verification, compare approved feature requirements with actual datasets and test a reduced-data version for utility would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs

D:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Question 219

A team considers this AI use: A recruitment platform provides rejection scores without checking obligations for human review and candidate requests. Rather than applying the relevant control, it will tell rejected applicants that model decisions can never be reviewed because the system is automated. What can go wrong?

  1. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  2. People's data may support decisions they were never reasonably told to expect.
  3. People may lose legally required recourse or receive no usable explanation of a consequential outcome.
  4. The organization may build a use that cannot lawfully operate in its target context.

Explanation

Correct Answer

C:

For the risk comparison, people may lose legally required recourse or receive no usable explanation of a consequential outcome. The shortcut overlooks the central safeguard: design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome

D:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: People may lose legally required recourse or receive no usable explanation of a consequential outcome.

Question 220

Choose the intervention and validation that fit this business problem: A recruitment platform provides rejection scores without checking obligations for human review and candidate requests.

  1. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.
  2. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  3. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  4. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Explanation

Correct Answer

D:

For the action and verification, design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs. Together, they implement and test the intended safeguard. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs

B:

This pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

C:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. The supporting evidence is: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Question 221

Consider the following proposed AI use: A language startup collects subscription-only news articles for model tuning without reading licenses. Which response best addresses the governance gap?

  1. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  2. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.
  3. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  4. Identify and document the applicable lawful basis before reusing personal information for model development.

Explanation

Correct Answer

A:

Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose. The practical acceptance evidence is: inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Incorrect Answers

B:

The alternative—review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

C:

The alternative—include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

D:

The alternative—identify and document the applicable lawful basis before reusing personal information for model development—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Question 222

A language startup collects subscription-only news articles for model tuning without reading licenses. Why is the recommended governance approach appropriate?

  1. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  2. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  3. Existing product-safety and liability principles can still matter when AI influences the operation of a product.
  4. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Explanation

Correct Answer

B:

Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose. That is why the case calls for this action: document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Incorrect Answers

A:

That explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

C:

The alternative—existing product-safety and liability principles can still matter when AI influences the operation of a product—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

D:

The alternative—a law may require demonstrable compliance; good intentions are not a substitute for reproducible records—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Question 223

The organization has implemented a control for this case: A language startup collects subscription-only news articles for model tuning without reading licenses. Which result would best validate it?

  1. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Compare published promises, disclaimers and actual tested limitations before approval.
  3. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  4. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Explanation

Correct Answer

C:

Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Incorrect Answers

A:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

B:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

D:

Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Question 224

The organization has this problem: A language startup collects subscription-only news articles for model tuning without reading licenses. It takes a shortcut: Scrape paid publications for training because the articles can be viewed in a browser. What is the principal risk?

  1. Customers may make costly decisions based on unsupported performance claims.
  2. A dangerous design defect or inadequate warning may go unaddressed.
  3. Key information and risk controls may be missing across the model supply chain.
  4. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.

Explanation

Correct Answer

D:

The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes. The shortcut overlooks the central safeguard: document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.

B:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.

C:

This consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.

Question 225

A language startup collects subscription-only news articles for model tuning without reading licenses. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Evidence: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  2. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Explanation

Correct Answer

A:

Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials. Together, they implement and test the intended safeguard. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Incorrect Answers

B:

This pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

C:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

D:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Question 226

A music metadata company proposes training on licensed editorial reviews outside contractual use terms. Which governance action most directly addresses the need in this situation?

  1. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.
  2. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  3. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.
  4. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Explanation

Correct Answer

B:

For the decision, document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose. The practical acceptance evidence is: inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials

Incorrect Answers

A:

For the decision, the alternative—review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model

C:

The alternative—review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

D:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Question 227

A music metadata company proposes training on licensed editorial reviews outside contractual use terms. Which explanation best connects the risk to the proposed governance action?

  1. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.
  2. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.
  3. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  4. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Explanation

Correct Answer

C:

For the technical explanation, public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose. That is why the case calls for this action: document rights and restrictions for copyrighted or licensed sources before training or adapting a model

Incorrect Answers

A:

That explains why a team might check the provider's contract and relevant law for permitted commercial use and responsibility for generated content, but it does not justify the decision required by the case. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

B:

The alternative—deceptive descriptions of model capabilities can create legal risk even when the underlying product functions—is not the most direct response to this case. The matching action is to document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

D:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Question 228

A music metadata company proposes training on licensed editorial reviews outside contractual use terms. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Compare published promises, disclaimers and actual tested limitations before approval.
  3. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  4. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Explanation

Correct Answer

D:

For the verification, inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose

Incorrect Answers

A:

For the verification, review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials

B:

For the verification, compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials

C:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Question 229

A music metadata company proposes training on licensed editorial reviews outside contractual use terms. The team instead proposes: Scrape paid publications for training because the articles can be viewed in a browser. Which failure should the reviewer expect most directly?

  1. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.
  2. Customers may make costly decisions based on unsupported performance claims.
  3. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.
  4. Key information and risk controls may be missing across the model supply chain.

Explanation

Correct Answer

A:

For the risk comparison, the training corpus may violate licensing limits or introduce avoidable intellectual-property disputes. The shortcut overlooks the central safeguard: document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes

C:

This consequence follows more directly from another shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes

Question 230

Choose the intervention and validation that fit this business problem: A music metadata company proposes training on licensed editorial reviews outside contractual use terms.

  1. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.
  2. Action: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Evidence: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  3. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.

Explanation

Correct Answer

B:

For the action and verification, document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials. Together, they implement and test the intended safeguard. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose

Incorrect Answers

A:

This pair supports another governance intervention: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

C:

For the action and verification, this pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials

D:

This pair supports another governance intervention: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For the actual case, the required action is: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. The supporting evidence is: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Question 231

A travel agency wants to print AI-generated graphics on merchandise without checking vendor terms. Which step would provide the most targeted control for the problem described?

  1. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.
  2. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  3. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.
  4. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.

Explanation

Correct Answer

C:

Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership. The practical acceptance evidence is: review license terms and approval records for representative public-facing generated materials.

Incorrect Answers

A:

The alternative—review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

B:

The alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

D:

The alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Question 232

A team reviewing the following situation asks for the most relevant governance principle: A travel agency wants to print AI-generated graphics on merchandise without checking vendor terms. Which principle best explains the required control?

  1. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  2. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  3. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.
  4. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Explanation

Correct Answer

D:

Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership. That is why the case calls for this action: check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Incorrect Answers

A:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

B:

That explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

C:

That explains why a team might design a process to identify relevant data subject requests and assess applicable automated-decision safeguards, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Question 233

A travel agency wants to print AI-generated graphics on merchandise without checking vendor terms. Which recorded test or observation most directly demonstrates an effective response?

  1. Review license terms and approval records for representative public-facing generated materials.
  2. Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Check the legal classification and an obligation-by-obligation evidence matrix before release.

Explanation

Correct Answer

A:

Review license terms and approval records for representative public-facing generated materials. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Incorrect Answers

B:

Evaluate selection and error rates across legally relevant groups, and document response to disparities would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials.

C:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials.

D:

Check the legal classification and an obligation-by-obligation evidence matrix before release would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials.

Question 234

A travel agency wants to print AI-generated graphics on merchandise without checking vendor terms. The team instead proposes: Assume every generated image is automatically free of third-party rights or usage limitations. Which failure should the reviewer expect most directly?

  1. Customers may make costly decisions based on unsupported performance claims.
  2. Public outputs can create contractual or infringement exposure even without intentional copying.
  3. A dangerous design defect or inadequate warning may go unaddressed.
  4. Technical scores do not create decision ownership or a risk governance structure.

Explanation

Correct Answer

B:

Public outputs can create contractual or infringement exposure even without intentional copying. The shortcut overlooks the central safeguard: check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying.

C:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying.

D:

This consequence follows more directly from another shortcut: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying.

Question 235

The governance team needs both a decision and a way to verify it for this case: A travel agency wants to print AI-generated graphics on merchandise without checking vendor terms. Which pair is best?

  1. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.
  4. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.

Explanation

Correct Answer

C:

Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Review license terms and approval records for representative public-facing generated materials. Together, they implement and test the intended safeguard. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Incorrect Answers

A:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials.

B:

This pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials.

D:

This pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials.

Question 236

Consider the following proposed AI use: A publisher plans to sell generated book illustrations and omits rights and similarity review. Which response best addresses the governance gap?

  1. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  2. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.
  3. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  4. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Explanation

Correct Answer

D:

For the decision, check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership. The practical acceptance evidence is: review license terms and approval records for representative public-facing generated materials

Incorrect Answers

A:

For the decision, the alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content

B:

The alternative—review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

C:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—is not the most direct response to this case. The matching action is to check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Question 237

A team reviewing the following situation asks for the most relevant governance principle: A publisher plans to sell generated book illustrations and omits rights and similarity review. Which principle best explains the required control?

  1. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.
  2. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  3. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  4. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Explanation

Correct Answer

A:

For the technical explanation, rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership. That is why the case calls for this action: check the provider's contract and relevant law for permitted commercial use and responsibility for generated content

Incorrect Answers

B:

For the technical explanation, that explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership

C:

For the technical explanation, that explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership

D:

That explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Question 238

An AI governance reviewer examines this situation: A publisher plans to sell generated book illustrations and omits rights and similarity review. Which acceptance check is most decisive?

  1. Compare published promises, disclaimers and actual tested limitations before approval.
  2. Review license terms and approval records for representative public-facing generated materials.
  3. Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  4. Record the original notice and collection purpose alongside the proposed use and compatibility determination.

Explanation

Correct Answer

B:

For the verification, review license terms and approval records for representative public-facing generated materials. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership

Incorrect Answers

A:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials.

C:

For the verification, evaluate selection and error rates across legally relevant groups, and document response to disparities would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials

D:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to review license terms and approval records for representative public-facing generated materials.

Question 239

For this case—A publisher plans to sell generated book illustrations and omits rights and similarity review—staff choose the shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. Which consequence best explains why this is inadequate?

  1. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.
  2. A dangerous design defect or inadequate warning may go unaddressed.
  3. Public outputs can create contractual or infringement exposure even without intentional copying.
  4. An unauthorized new processing purpose can be hidden behind existing access rights.

Explanation

Correct Answer

C:

For the risk comparison, public outputs can create contractual or infringement exposure even without intentional copying. The shortcut overlooks the central safeguard: check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying

D:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Public outputs can create contractual or infringement exposure even without intentional copying.

Question 240

A publisher plans to sell generated book illustrations and omits rights and similarity review. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  2. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  3. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  4. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.

Explanation

Correct Answer

D:

For the action and verification, check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Review license terms and approval records for representative public-facing generated materials. Together, they implement and test the intended safeguard. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials

B:

For the action and verification, this pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials

C:

This pair supports another governance intervention: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For the actual case, the required action is: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. The supporting evidence is: Review license terms and approval records for representative public-facing generated materials.

Question 241

A logistics company scores forklift applicants using prior job tenure correlated with unequal access. Which step would provide the most targeted control for the problem described?

  1. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.
  2. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.
  3. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  4. Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Explanation

Correct Answer

A:

Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The practical acceptance evidence is: evaluate selection and error rates across legally relevant groups, and document response to disparities.

Incorrect Answers

B:

This would help address a separate governance task: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

C:

This would help address a separate governance task: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

D:

This would help address a separate governance task: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

Question 242

A logistics company scores forklift applicants using prior job tenure correlated with unequal access. Why is the recommended governance approach appropriate?

  1. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  2. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  3. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.
  4. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Explanation

Correct Answer

B:

Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. That is why the case calls for this action: review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

Incorrect Answers

A:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

C:

That explains why a team might review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

D:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Question 243

The organization has implemented a control for this case: A logistics company scores forklift applicants using prior job tenure correlated with unequal access. Which result would best validate it?

  1. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  3. Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  4. Review provider technical documentation and downstream integration requirements under the relevant regime.

Explanation

Correct Answer

C:

Evaluate selection and error rates across legally relevant groups, and document response to disparities. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Incorrect Answers

A:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

B:

Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

D:

Review provider technical documentation and downstream integration requirements under the relevant regime would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

Question 244

The organization has this problem: A logistics company scores forklift applicants using prior job tenure correlated with unequal access. It takes a shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. What is the principal risk?

  1. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.
  2. A dangerous design defect or inadequate warning may go unaddressed.
  3. The organization can overlook use context and fail to act on significant measured risks.
  4. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

Explanation

Correct Answer

D:

Proxy effects can cause systematically unequal opportunities even with no protected attribute field. The shortcut overlooks the central safeguard: review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Scrape paid publications for training because the articles can be viewed in a browser. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

B:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

C:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

Question 245

The governance team needs both a decision and a way to verify it for this case: A logistics company scores forklift applicants using prior job tenure correlated with unequal access. Which pair is best?

  1. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  2. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.
  3. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Explanation

Correct Answer

A:

Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evaluate selection and error rates across legally relevant groups, and document response to disparities. Together, they implement and test the intended safeguard. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Incorrect Answers

B:

This pair supports another governance intervention: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities.

C:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities.

D:

This pair supports another governance intervention: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities.

Question 246

A software company deploys interview ranking without testing whether comparable candidates receive different outcomes. Which step would provide the most targeted control for the problem described?

  1. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  2. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.
  3. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  4. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.

Explanation

Correct Answer

B:

For the decision, review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The practical acceptance evidence is: evaluate selection and error rates across legally relevant groups, and document response to disparities

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements

C:

This would help address a separate governance task: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

D:

This would help address a separate governance task: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. For this case, however, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. The more direct step is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.

Question 247

A team reviewing the following situation asks for the most relevant governance principle: A software company deploys interview ranking without testing whether comparable candidates receive different outcomes. Which principle best explains the required control?

  1. Existing product-safety and liability principles can still matter when AI influences the operation of a product.
  2. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.
  3. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  4. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Explanation

Correct Answer

C:

For the technical explanation, using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias. That is why the case calls for this action: review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements

Incorrect Answers

A:

That explains why a team might include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

B:

That explains why a team might check the provider's contract and relevant law for permitted commercial use and responsibility for generated content, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

D:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.

Question 248

A software company deploys interview ranking without testing whether comparable candidates receive different outcomes. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Compare published promises, disclaimers and actual tested limitations before approval.
  2. Review license terms and approval records for representative public-facing generated materials.
  3. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Evaluate selection and error rates across legally relevant groups, and document response to disparities.

Explanation

Correct Answer

D:

For the verification, evaluate selection and error rates across legally relevant groups, and document response to disparities. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias

Incorrect Answers

A:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

B:

Review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

C:

Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime would test a different outcome. The directly relevant evidence is to evaluate selection and error rates across legally relevant groups, and document response to disparities.

Question 249

For this case—A software company deploys interview ranking without testing whether comparable candidates receive different outcomes—staff choose the shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. Which consequence best explains why this is inadequate?

  1. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.
  2. Customers may make costly decisions based on unsupported performance claims.
  3. Public outputs can create contractual or infringement exposure even without intentional copying.
  4. An unauthorized new processing purpose can be hidden behind existing access rights.

Explanation

Correct Answer

A:

For the risk comparison, proxy effects can cause systematically unequal opportunities even with no protected attribute field. The shortcut overlooks the central safeguard: review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

C:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

D:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

Question 250

A software company deploys interview ranking without testing whether comparable candidates receive different outcomes. Which action and evidence pair most directly addresses the situation?

  1. Action: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Evidence: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  2. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.

Explanation

Correct Answer

B:

For the action and verification, review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evaluate selection and error rates across legally relevant groups, and document response to disparities. Together, they implement and test the intended safeguard. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias

Incorrect Answers

A:

This pair supports another governance intervention: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities.

C:

For the action and verification, this pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities

D:

This pair supports another governance intervention: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For the actual case, the required action is: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. The supporting evidence is: Evaluate selection and error rates across legally relevant groups, and document response to disparities.

Question 251

Consider the following proposed AI use: A tax app promotes an assistant as guaranteed to find every deduction despite known errors. Which response best addresses the governance gap?

  1. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  2. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.
  3. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.
  4. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

Explanation

Correct Answer

C:

Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The practical acceptance evidence is: compare published promises, disclaimers and actual tested limitations before approval.

Incorrect Answers

A:

This would help address a separate governance task: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For this case, however, deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The more direct step is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

B:

This would help address a separate governance task: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. For this case, however, deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The more direct step is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

D:

The alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review—would not resolve the stated issue. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Question 252

A team reviewing the following situation asks for the most relevant governance principle: A tax app promotes an assistant as guaranteed to find every deduction despite known errors. Which principle best explains the required control?

  1. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  2. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  3. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  4. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Explanation

Correct Answer

D:

Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. That is why the case calls for this action: review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

Incorrect Answers

A:

That explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

B:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

C:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Question 253

A tax app promotes an assistant as guaranteed to find every deduction despite known errors. Which recorded test or observation most directly demonstrates an effective response?

  1. Compare published promises, disclaimers and actual tested limitations before approval.
  2. Review license terms and approval records for representative public-facing generated materials.
  3. Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  4. Map internal policy commitments to concrete controls and evidence against the OECD principles.

Explanation

Correct Answer

A:

Compare published promises, disclaimers and actual tested limitations before approval. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Incorrect Answers

B:

Review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval.

C:

Evaluate selection and error rates across legally relevant groups, and document response to disparities would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval.

D:

Map internal policy commitments to concrete controls and evidence against the OECD principles would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval.

Question 254

For this case—A tax app promotes an assistant as guaranteed to find every deduction despite known errors—staff choose the shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. Which consequence best explains why this is inadequate?

  1. Public outputs can create contractual or infringement exposure even without intentional copying.
  2. Customers may make costly decisions based on unsupported performance claims.
  3. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.
  4. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

Explanation

Correct Answer

B:

Customers may make costly decisions based on unsupported performance claims. The shortcut overlooks the central safeguard: review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

C:

This consequence follows more directly from another shortcut: Scrape paid publications for training because the articles can be viewed in a browser. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

D:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

Question 255

Choose the intervention and validation that fit this business problem: A tax app promotes an assistant as guaranteed to find every deduction despite known errors.

  1. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.
  2. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.
  4. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

C:

Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Compare published promises, disclaimers and actual tested limitations before approval. Together, they implement and test the intended safeguard. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Incorrect Answers

A:

This pair supports another governance intervention: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval.

B:

This pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval.

D:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval.

Question 256

An AI project team faces this situation: A wellness platform markets an AI recommendation as clinically validated without supporting evidence. What should the team do first to manage the central issue?

  1. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  2. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  3. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  4. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

Explanation

Correct Answer

D:

For the decision, review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The practical acceptance evidence is: compare published promises, disclaimers and actual tested limitations before approval

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For this case, however, deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The more direct step is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use

B:

This would help address a separate governance task: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For this case, however, deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. The more direct step is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

C:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—would not resolve the stated issue. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Question 257

The organization chooses to review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use for this situation: A wellness platform markets an AI recommendation as clinically validated without supporting evidence. Which rationale best supports that choice?

  1. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.
  2. Existing product-safety and liability principles can still matter when AI influences the operation of a product.
  3. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  4. Reusing information for an unrelated objective can create a new privacy obligation despite good technical safeguards.

Explanation

Correct Answer

A:

For the technical explanation, deceptive descriptions of model capabilities can create legal risk even when the underlying product functions. That is why the case calls for this action: review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use

Incorrect Answers

B:

That explains why a team might include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

C:

For the technical explanation, that explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions

D:

That explains why a team might test whether a proposed new model purpose is compatible with the reason personal information was originally obtained, but it does not justify the decision required by the case. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Question 258

A wellness platform markets an AI recommendation as clinically validated without supporting evidence. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Review license terms and approval records for representative public-facing generated materials.
  2. Compare published promises, disclaimers and actual tested limitations before approval.
  3. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  4. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Explanation

Correct Answer

B:

For the verification, compare published promises, disclaimers and actual tested limitations before approval. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions

Incorrect Answers

A:

For the verification, review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval

C:

Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval.

D:

Trace a selected use case from contextual mapping to measured evidence and a documented risk response would test a different outcome. The directly relevant evidence is to compare published promises, disclaimers and actual tested limitations before approval.

Question 259

For this case—A wellness platform markets an AI recommendation as clinically validated without supporting evidence—staff choose the shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. Which consequence best explains why this is inadequate?

  1. A dangerous design defect or inadequate warning may go unaddressed.
  2. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.
  3. Customers may make costly decisions based on unsupported performance claims.
  4. People's data may support decisions they were never reasonably told to expect.

Explanation

Correct Answer

C:

For the risk comparison, customers may make costly decisions based on unsupported performance claims. The shortcut overlooks the central safeguard: review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

B:

This consequence follows more directly from another shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

D:

This consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: Customers may make costly decisions based on unsupported performance claims.

Question 260

The governance team needs both a decision and a way to verify it for this case: A wellness platform markets an AI recommendation as clinically validated without supporting evidence. Which pair is best?

  1. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Action: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. Evidence: Review license terms and approval records for representative public-facing generated materials.
  3. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  4. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.

Explanation

Correct Answer

D:

For the action and verification, review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Compare published promises, disclaimers and actual tested limitations before approval. Together, they implement and test the intended safeguard. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions

Incorrect Answers

A:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval.

B:

For the action and verification, this pair supports another governance intervention: Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval

C:

This pair supports another governance intervention: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For the actual case, the required action is: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. The supporting evidence is: Compare published promises, disclaimers and actual tested limitations before approval.

Question 261

Consider the following proposed AI use: An industrial robot uses vision-based pedestrian detection without evaluating model failure cases. Which response best addresses the governance gap?

  1. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  2. Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements.
  3. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  4. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.

Explanation

Correct Answer

A:

Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Existing product-safety and liability principles can still matter when AI influences the operation of a product. The practical acceptance evidence is: review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Incorrect Answers

B:

This would help address a separate governance task: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For this case, however, existing product-safety and liability principles can still matter when AI influences the operation of a product. The more direct step is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.

C:

The alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—would not resolve the stated issue. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

D:

The alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review—would not resolve the stated issue. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Question 262

An industrial robot uses vision-based pedestrian detection without evaluating model failure cases. Which explanation best connects the risk to the proposed governance action?

  1. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.
  2. Existing product-safety and liability principles can still matter when AI influences the operation of a product.
  3. Using a neutral-looking feature set does not prove that employment decisions are free from unlawful bias.
  4. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Explanation

Correct Answer

B:

Existing product-safety and liability principles can still matter when AI influences the operation of a product. That is why the case calls for this action: include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.

Incorrect Answers

A:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

C:

That explains why a team might review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

D:

That explains why a team might design a process to identify relevant data subject requests and assess applicable automated-decision safeguards, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Question 263

An AI governance reviewer examines this situation: An industrial robot uses vision-based pedestrian detection without evaluating model failure cases. Which acceptance check is most decisive?

  1. Compare published promises, disclaimers and actual tested limitations before approval.
  2. Review license terms and approval records for representative public-facing generated materials.
  3. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Test representative user journeys for the required AI notices and correct timing.

Explanation

Correct Answer

C:

Review documented hazards, foreseeable misuse, safety testing and corrective action ownership. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Incorrect Answers

A:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

B:

Review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

D:

Test representative user journeys for the required AI notices and correct timing would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Question 264

An industrial robot uses vision-based pedestrian detection without evaluating model failure cases. The team instead proposes: Treat a safety-critical model as outside product controls merely because it runs in software. Which failure should the reviewer expect most directly?

  1. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.
  2. Public outputs can create contractual or infringement exposure even without intentional copying.
  3. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  4. A dangerous design defect or inadequate warning may go unaddressed.

Explanation

Correct Answer

D:

A dangerous design defect or inadequate warning may go unaddressed. The shortcut overlooks the central safeguard: include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Scrape paid publications for training because the articles can be viewed in a browser. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed.

B:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed.

C:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed.

Question 265

An industrial robot uses vision-based pedestrian detection without evaluating model failure cases. Which action and evidence pair most directly addresses the situation?

  1. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  2. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  3. Action: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Evidence: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  4. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

A:

Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership. Together, they implement and test the intended safeguard. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Incorrect Answers

B:

This pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

C:

This pair supports another governance intervention: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

D:

This pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Question 266

Consider the following proposed AI use: A home appliance adds voice-controlled heating and skips investigation of unsafe activation commands. Which response best addresses the governance gap?

  1. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.
  2. Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.
  3. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.
  4. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Explanation

Correct Answer

B:

For the decision, include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Existing product-safety and liability principles can still matter when AI influences the operation of a product. The practical acceptance evidence is: review documented hazards, foreseeable misuse, safety testing and corrective action ownership

Incorrect Answers

A:

The alternative—check the provider's contract and relevant law for permitted commercial use and responsibility for generated content—would not resolve the stated issue. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

C:

For the decision, the alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—would not resolve the stated issue. Existing product-safety and liability principles can still matter when AI influences the operation of a product

D:

This would help address a separate governance task: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. For this case, however, existing product-safety and liability principles can still matter when AI influences the operation of a product. The more direct step is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment.

Question 267

A home appliance adds voice-controlled heating and skips investigation of unsafe activation commands. Why is the recommended governance approach appropriate?

  1. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.
  2. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.
  3. Existing product-safety and liability principles can still matter when AI influences the operation of a product.
  4. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Explanation

Correct Answer

C:

For the technical explanation, existing product-safety and liability principles can still matter when AI influences the operation of a product. That is why the case calls for this action: include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment

Incorrect Answers

A:

That explains why a team might check the provider's contract and relevant law for permitted commercial use and responsibility for generated content, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

B:

That explains why a team might review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

D:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. Existing product-safety and liability principles can still matter when AI influences the operation of a product.

Question 268

A home appliance adds voice-controlled heating and skips investigation of unsafe activation commands. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Review license terms and approval records for representative public-facing generated materials.
  2. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  3. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  4. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

D:

For the verification, review documented hazards, foreseeable misuse, safety testing and corrective action ownership. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Existing product-safety and liability principles can still matter when AI influences the operation of a product

Incorrect Answers

A:

For the verification, review license terms and approval records for representative public-facing generated materials would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership

B:

Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

C:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Question 269

The organization has this problem: A home appliance adds voice-controlled heating and skips investigation of unsafe activation commands. It takes a shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. What is the principal risk?

  1. A dangerous design defect or inadequate warning may go unaddressed.
  2. Public outputs can create contractual or infringement exposure even without intentional copying.
  3. The training corpus may violate licensing limits or introduce avoidable intellectual-property disputes.
  4. The organization can overlook use context and fail to act on significant measured risks.

Explanation

Correct Answer

A:

For the risk comparison, a dangerous design defect or inadequate warning may go unaddressed. The shortcut overlooks the central safeguard: include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Existing product-safety and liability principles can still matter when AI influences the operation of a product

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed

C:

For the risk comparison, this consequence follows more directly from another shortcut: Scrape paid publications for training because the articles can be viewed in a browser. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed

D:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: A dangerous design defect or inadequate warning may go unaddressed.

Question 270

The governance team needs both a decision and a way to verify it for this case: A home appliance adds voice-controlled heating and skips investigation of unsafe activation commands. Which pair is best?

  1. Action: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. Evidence: Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.
  2. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  3. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.
  4. Action: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. Evidence: Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Explanation

Correct Answer

B:

For the action and verification, include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership. Together, they implement and test the intended safeguard. Existing product-safety and liability principles can still matter when AI influences the operation of a product

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Document rights and restrictions for copyrighted or licensed sources before training or adapting a model. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership

C:

This pair supports another governance intervention: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

D:

This pair supports another governance intervention: Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient. For the actual case, the required action is: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. The supporting evidence is: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Question 271

A school vendor proposes evaluating children by inferred emotional state under a risk-based AI regime. Which governance action most directly addresses the need in this situation?

  1. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  2. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.
  3. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.
  4. Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use.

Explanation

Correct Answer

C:

Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition. The practical acceptance evidence is: record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Incorrect Answers

A:

The alternative—identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

B:

The alternative—differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

D:

The alternative—review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 272

A school vendor proposes evaluating children by inferred emotional state under a risk-based AI regime. Why is the recommended governance approach appropriate?

  1. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.
  2. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  3. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  4. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

Explanation

Correct Answer

D:

Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition. That is why the case calls for this action: conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Incorrect Answers

A:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

B:

The alternative—a law may require demonstrable compliance; good intentions are not a substitute for reproducible records—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

C:

The alternative—different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 273

A school vendor proposes evaluating children by inferred emotional state under a risk-based AI regime. Which recorded test or observation most directly demonstrates an effective response?

  1. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  4. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Explanation

Correct Answer

A:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

Incorrect Answers

B:

This evidence could test a different control, such as: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. It does not establish the safeguard needed here. A decisive check is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

C:

The alternative—simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

D:

The alternative—trace a selected use case from contextual mapping to measured evidence and a documented risk response—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 274

The organization has this problem: A school vendor proposes evaluating children by inferred emotional state under a risk-based AI regime. It takes a shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. What is the principal risk?

  1. Key information and risk controls may be missing across the model supply chain.
  2. The organization may build a use that cannot lawfully operate in its target context.
  3. People may be misled about interacting with AI or receiving AI-generated material.
  4. An unauthorized new processing purpose can be hidden behind existing access rights.

Explanation

Correct Answer

B:

The organization may build a use that cannot lawfully operate in its target context. The shortcut overlooks the central safeguard: conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context.

C:

This consequence follows more directly from another shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context.

D:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context.

Question 275

A school vendor proposes evaluating children by inferred emotional state under a risk-based AI regime. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  2. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  4. Action: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. Evidence: Record the original notice and collection purpose alongside the proposed use and compatibility determination.

Explanation

Correct Answer

C:

Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition. Together, they implement and test the intended safeguard. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

Incorrect Answers

A:

This pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

B:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

D:

This pair supports another governance intervention: Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Question 276

Consider the following proposed AI use: An employer proposes covert behavioral manipulation using a workplace assistant in a regulated market. Which response best addresses the governance gap?

  1. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  2. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.
  3. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.
  4. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Explanation

Correct Answer

D:

For the decision, conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition. The practical acceptance evidence is: record the affected people, purpose, jurisdiction, current legal classification and legal disposition

Incorrect Answers

A:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

B:

For the decision, the alternative—differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development

C:

The alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 277

An employer proposes covert behavioral manipulation using a workplace assistant in a regulated market. Why is the recommended governance approach appropriate?

  1. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.
  2. General-purpose model duties and downstream application controls are related but not interchangeable.
  3. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  4. Collecting more personal data than required increases exposure without automatically improving the model.

Explanation

Correct Answer

A:

For the technical explanation, some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition. That is why the case calls for this action: conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development

Incorrect Answers

B:

That explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

C:

That explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.

D:

The alternative—collecting more personal data than required increases exposure without automatically improving the model—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 278

The organization has implemented a control for this case: An employer proposes covert behavioral manipulation using a workplace assistant in a regulated market. Which result would best validate it?

  1. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  3. Test representative user journeys for the required AI notices and correct timing.
  4. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

B:

For the verification, record the affected people, purpose, jurisdiction, current legal classification and legal disposition. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition

Incorrect Answers

A:

For the verification, the alternative—simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development

C:

The alternative—test representative user journeys for the required AI notices and correct timing—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

D:

The alternative—review documented hazards, foreseeable misuse, safety testing and corrective action ownership—is not the most direct response to this case. The matching action is to conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.

Question 279

For this case—An employer proposes covert behavioral manipulation using a workplace assistant in a regulated market—staff choose the shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. Which consequence best explains why this is inadequate?

  1. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  2. Key information and risk controls may be missing across the model supply chain.
  3. The organization may build a use that cannot lawfully operate in its target context.
  4. A dangerous design defect or inadequate warning may go unaddressed.

Explanation

Correct Answer

C:

For the risk comparison, the organization may build a use that cannot lawfully operate in its target context. The shortcut overlooks the central safeguard: conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context

D:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: The organization may build a use that cannot lawfully operate in its target context.

Question 280

An employer proposes covert behavioral manipulation using a workplace assistant in a regulated market. Which action and evidence pair most directly addresses the situation?

  1. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  2. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.
  3. Action: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. Evidence: Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.
  4. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Explanation

Correct Answer

D:

For the action and verification, conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition. Together, they implement and test the intended safeguard. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition

B:

This pair supports another governance intervention: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

C:

This pair supports another governance intervention: Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards. For the actual case, the required action is: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. The supporting evidence is: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.

Question 281

An EU-facing recruiting vendor classifies a selection system as high-risk and asks which evidence to assemble. Which governance action most directly addresses the need in this situation?

  1. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  2. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  3. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  4. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

Explanation

Correct Answer

A:

For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role. The practical acceptance evidence is: check the legal classification and an obligation-by-obligation evidence matrix before release.

Incorrect Answers

B:

The alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

C:

The alternative—identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

D:

The alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Question 282

A team reviewing the following situation asks for the most relevant governance principle: An EU-facing recruiting vendor classifies a selection system as high-risk and asks which evidence to assemble. Which principle best explains the required control?

  1. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  2. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  3. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  4. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.

Explanation

Correct Answer

B:

A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role. That is why the case calls for this action: for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.

Incorrect Answers

A:

That explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

C:

That explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

D:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Question 283

An AI governance reviewer examines this situation: An EU-facing recruiting vendor classifies a selection system as high-risk and asks which evidence to assemble. Which acceptance check is most decisive?

  1. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  3. Check the legal classification and an obligation-by-obligation evidence matrix before release.
  4. Compare approved feature requirements with actual datasets and test a reduced-data version for utility.

Explanation

Correct Answer

C:

Check the legal classification and an obligation-by-obligation evidence matrix before release. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Incorrect Answers

A:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release.

B:

Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release.

D:

Compare approved feature requirements with actual datasets and test a reduced-data version for utility would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release.

Question 284

A team considers this AI use: An EU-facing recruiting vendor classifies a selection system as high-risk and asks which evidence to assemble. Rather than applying the relevant control, it will attach a generic responsible-AI statement instead of testing specific high-risk requirements. What can go wrong?

  1. People may be misled about interacting with AI or receiving AI-generated material.
  2. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  3. People's data may support decisions they were never reasonably told to expect.
  4. Required assurance, documentation or human oversight may be missing at a legal checkpoint.

Explanation

Correct Answer

D:

Required assurance, documentation or human oversight may be missing at a legal checkpoint. The shortcut overlooks the central safeguard: for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

B:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

C:

This consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

Question 285

Choose the intervention and validation that fit this business problem: An EU-facing recruiting vendor classifies a selection system as high-risk and asks which evidence to assemble.

  1. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.
  2. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  4. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

A:

For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Check the legal classification and an obligation-by-obligation evidence matrix before release. Together, they implement and test the intended safeguard. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Incorrect Answers

B:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release.

C:

This pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release.

D:

This pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release.

Question 286

Consider the following proposed AI use: A supplier of regulated qualification assessments identifies heightened obligations for its scoring service. Which response best addresses the governance gap?

  1. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  2. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  3. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  4. Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review.

Explanation

Correct Answer

B:

For the decision, for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role. The practical acceptance evidence is: check the legal classification and an obligation-by-obligation evidence matrix before release

Incorrect Answers

A:

For the decision, the alternative—identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role

C:

For the decision, the alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role

D:

The alternative—escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review—would not resolve the stated issue. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Question 287

The organization chooses to for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls for this situation: A supplier of regulated qualification assessments identifies heightened obligations for its scoring service. Which rationale best supports that choice?

  1. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.
  2. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.
  3. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  4. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Explanation

Correct Answer

C:

For the technical explanation, a high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role. That is why the case calls for this action: for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls

Incorrect Answers

A:

That explains why a team might conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

B:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

D:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.

Question 288

An AI governance reviewer examines this situation: A supplier of regulated qualification assessments identifies heightened obligations for its scoring service. Which acceptance check is most decisive?

  1. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  3. Record the original notice and collection purpose alongside the proposed use and compatibility determination.
  4. Check the legal classification and an obligation-by-obligation evidence matrix before release.

Explanation

Correct Answer

D:

For the verification, check the legal classification and an obligation-by-obligation evidence matrix before release. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role

Incorrect Answers

A:

For the verification, record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release

B:

For the verification, simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release

C:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to check the legal classification and an obligation-by-obligation evidence matrix before release.

Question 289

A supplier of regulated qualification assessments identifies heightened obligations for its scoring service. The team instead proposes: Attach a generic responsible-AI statement instead of testing specific high-risk requirements. Which failure should the reviewer expect most directly?

  1. Required assurance, documentation or human oversight may be missing at a legal checkpoint.
  2. The organization may build a use that cannot lawfully operate in its target context.
  3. Key information and risk controls may be missing across the model supply chain.
  4. Important management or impact-assessment activities may be omitted under the wrong standard.

Explanation

Correct Answer

A:

For the risk comparison, required assurance, documentation or human oversight may be missing at a legal checkpoint. The shortcut overlooks the central safeguard: for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

C:

This consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

D:

This consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: Required assurance, documentation or human oversight may be missing at a legal checkpoint.

Question 290

Choose the intervention and validation that fit this business problem: A supplier of regulated qualification assessments identifies heightened obligations for its scoring service.

  1. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.
  3. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

B:

For the action and verification, for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Check the legal classification and an obligation-by-obligation evidence matrix before release. Together, they implement and test the intended safeguard. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release

C:

For the action and verification, this pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release

D:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. The supporting evidence is: Check the legal classification and an obligation-by-obligation evidence matrix before release.

Question 291

Consider the following proposed AI use: A retailer adds a synthetic spokesperson to its support channel without assessing disclosure duties. Which response best addresses the governance gap?

  1. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  2. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  3. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  4. Design a process to identify relevant data subject requests and assess applicable automated-decision safeguards.

Explanation

Correct Answer

C:

Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights. The practical acceptance evidence is: test representative user journeys for the required AI notices and correct timing.

Incorrect Answers

A:

The alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review—would not resolve the stated issue. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

B:

This would help address a separate governance task: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For this case, however, transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights. The more direct step is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

D:

The alternative—design a process to identify relevant data subject requests and assess applicable automated-decision safeguards—would not resolve the stated issue. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

Question 292

The organization chooses to identify whether an AI interaction or generated content requires specific disclosure in the applicable context for this situation: A retailer adds a synthetic spokesperson to its support channel without assessing disclosure duties. Which rationale best supports that choice?

  1. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.
  2. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  3. Collecting more personal data than required increases exposure without automatically improving the model.
  4. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Explanation

Correct Answer

D:

Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights. That is why the case calls for this action: identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

Incorrect Answers

A:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

B:

That explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

C:

That explains why a team might remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Question 293

An AI governance reviewer examines this situation: A retailer adds a synthetic spokesperson to its support channel without assessing disclosure duties. Which acceptance check is most decisive?

  1. Test representative user journeys for the required AI notices and correct timing.
  2. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  4. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Explanation

Correct Answer

A:

Test representative user journeys for the required AI notices and correct timing. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Incorrect Answers

B:

Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing.

C:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing.

D:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing.

Question 294

A team considers this AI use: A retailer adds a synthetic spokesperson to its support channel without assessing disclosure duties. Rather than applying the relevant control, it will hide the use of a customer-facing synthetic voice because the answer content is plausible. What can go wrong?

  1. Required assurance, documentation or human oversight may be missing at a legal checkpoint.
  2. People may be misled about interacting with AI or receiving AI-generated material.
  3. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  4. A dangerous design defect or inadequate warning may go unaddressed.

Explanation

Correct Answer

B:

People may be misled about interacting with AI or receiving AI-generated material. The shortcut overlooks the central safeguard: identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Attach a generic responsible-AI statement instead of testing specific high-risk requirements. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

C:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

D:

This consequence follows more directly from another shortcut: Treat a safety-critical model as outside product controls merely because it runs in software. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

Question 295

A retailer adds a synthetic spokesperson to its support channel without assessing disclosure duties. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  2. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  3. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.
  4. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Explanation

Correct Answer

C:

Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Test representative user journeys for the required AI notices and correct timing. Together, they implement and test the intended safeguard. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Incorrect Answers

A:

This pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing.

B:

This pair supports another governance intervention: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing.

D:

This pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing.

Question 296

A tourism board publishes AI-generated public-information audio without checking labeling rules. Which step would provide the most targeted control for the problem described?

  1. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  2. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.
  3. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  4. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

Explanation

Correct Answer

D:

For the decision, identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights. The practical acceptance evidence is: test representative user journeys for the required AI notices and correct timing

Incorrect Answers

A:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—would not resolve the stated issue. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

B:

The alternative—differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law—would not resolve the stated issue. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

C:

The alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—would not resolve the stated issue. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.

Question 297

A tourism board publishes AI-generated public-information audio without checking labeling rules. Why is the recommended governance approach appropriate?

  1. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  2. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  3. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  4. Rights in model outputs and third-party content depend on contract and applicable law; fluency is not evidence of ownership.

Explanation

Correct Answer

A:

For the technical explanation, transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights. That is why the case calls for this action: identify whether an AI interaction or generated content requires specific disclosure in the applicable context

Incorrect Answers

B:

That explains why a team might for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

C:

For the technical explanation, that explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights

D:

That explains why a team might check the provider's contract and relevant law for permitted commercial use and responsibility for generated content, but it does not justify the decision required by the case. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Question 298

A tourism board publishes AI-generated public-information audio without checking labeling rules. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Test representative user journeys for the required AI notices and correct timing.
  3. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

B:

For the verification, test representative user journeys for the required AI notices and correct timing. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights

Incorrect Answers

A:

Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing.

C:

For the verification, map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing

D:

Verify that the chosen reference matches the control objective and that the project cites the correct standard would test a different outcome. The directly relevant evidence is to test representative user journeys for the required AI notices and correct timing.

Question 299

The organization has this problem: A tourism board publishes AI-generated public-information audio without checking labeling rules. It takes a shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. What is the principal risk?

  1. Key information and risk controls may be missing across the model supply chain.
  2. A role change may leave mandatory obligations or documentation unassigned.
  3. People may be misled about interacting with AI or receiving AI-generated material.
  4. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

Explanation

Correct Answer

C:

For the risk comparison, people may be misled about interacting with AI or receiving AI-generated material. The shortcut overlooks the central safeguard: identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

B:

This consequence follows more directly from another shortcut: Assume every organization interacting with an AI system is legally only a passive end user. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

D:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: People may be misled about interacting with AI or receiving AI-generated material.

Question 300

Choose the intervention and validation that fit this business problem: A tourism board publishes AI-generated public-information audio without checking labeling rules.

  1. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  3. Action: Identify and document the applicable lawful basis before reusing personal information for model development. Evidence: Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  4. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.

Explanation

Correct Answer

D:

For the action and verification, identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Test representative user journeys for the required AI notices and correct timing. Together, they implement and test the intended safeguard. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights

Incorrect Answers

A:

This pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing.

B:

For the action and verification, this pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing

C:

This pair supports another governance intervention: Identify and document the applicable lawful basis before reusing personal information for model development. For the actual case, the required action is: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. The supporting evidence is: Test representative user journeys for the required AI notices and correct timing.

Question 301

A local software vendor builds customer assistants on an externally supplied foundation model. Which step would provide the most targeted control for the problem described?

  1. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  2. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  3. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  4. Identify and document the applicable lawful basis before reusing personal information for model development.

Explanation

Correct Answer

A:

Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. General-purpose model duties and downstream application controls are related but not interchangeable. The practical acceptance evidence is: review provider technical documentation and downstream integration requirements under the relevant regime.

Incorrect Answers

B:

This would help address a separate governance task: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

C:

This would help address a separate governance task: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

D:

This would help address a separate governance task: Identify and document the applicable lawful basis before reusing personal information for model development. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

Question 302

A local software vendor builds customer assistants on an externally supplied foundation model. Why is the recommended governance approach appropriate?

  1. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  2. General-purpose model duties and downstream application controls are related but not interchangeable.
  3. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  4. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Explanation

Correct Answer

B:

General-purpose model duties and downstream application controls are related but not interchangeable. That is why the case calls for this action: identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

Incorrect Answers

A:

That explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable.

C:

That explains why a team might for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable.

D:

That explains why a team might use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable.

Question 303

An AI governance reviewer examines this situation: A local software vendor builds customer assistants on an externally supplied foundation model. Which acceptance check is most decisive?

  1. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Review provider technical documentation and downstream integration requirements under the relevant regime.
  4. Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials.

Explanation

Correct Answer

C:

Review provider technical documentation and downstream integration requirements under the relevant regime. This verifies the necessary safeguard rather than merely showing that an AI project was completed. General-purpose model duties and downstream application controls are related but not interchangeable.

Incorrect Answers

A:

Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

B:

Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

D:

Inspect dataset-level source licenses, exceptions or permissions and restrict unverified materials would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

Question 304

For this case—A local software vendor builds customer assistants on an externally supplied foundation model—staff choose the shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. Which consequence best explains why this is inadequate?

  1. Required assurance, documentation or human oversight may be missing at a legal checkpoint.
  2. People may be misled about interacting with AI or receiving AI-generated material.
  3. An unauthorized new processing purpose can be hidden behind existing access rights.
  4. Key information and risk controls may be missing across the model supply chain.

Explanation

Correct Answer

D:

Key information and risk controls may be missing across the model supply chain. The shortcut overlooks the central safeguard: identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. General-purpose model duties and downstream application controls are related but not interchangeable.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Attach a generic responsible-AI statement instead of testing specific high-risk requirements. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain.

B:

This consequence follows more directly from another shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain.

C:

This consequence follows more directly from another shortcut: Assume a dataset may be trained on because staff already can open the underlying system. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain.

Question 305

Choose the intervention and validation that fit this business problem: A local software vendor builds customer assistants on an externally supplied foundation model.

  1. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  2. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.
  3. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.
  4. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

A:

Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Review provider technical documentation and downstream integration requirements under the relevant regime. Together, they implement and test the intended safeguard. General-purpose model duties and downstream application controls are related but not interchangeable.

Incorrect Answers

B:

This pair supports another governance intervention: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime.

C:

This pair supports another governance intervention: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime.

D:

This pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime.

Question 306

A cloud developer offers a general-purpose text model to multiple downstream application teams. Which step would provide the most targeted control for the problem described?

  1. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  2. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  3. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  4. Identify and document the applicable lawful basis before reusing personal information for model development.

Explanation

Correct Answer

B:

For the decision, identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. General-purpose model duties and downstream application controls are related but not interchangeable. The practical acceptance evidence is: review provider technical documentation and downstream integration requirements under the relevant regime

Incorrect Answers

A:

This would help address a separate governance task: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

C:

For the decision, this would help address a separate governance task: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities

D:

For the decision, this would help address a separate governance task: Identify and document the applicable lawful basis before reusing personal information for model development. For this case, however, general-purpose model duties and downstream application controls are related but not interchangeable. The more direct step is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities

Question 307

The organization chooses to identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities for this situation: A cloud developer offers a general-purpose text model to multiple downstream application teams. Which rationale best supports that choice?

  1. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  2. Some AI uses face outright restrictions; labeling a system innovative does not resolve an applicable prohibition.
  3. General-purpose model duties and downstream application controls are related but not interchangeable.
  4. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Explanation

Correct Answer

C:

For the technical explanation, general-purpose model duties and downstream application controls are related but not interchangeable. That is why the case calls for this action: identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities

Incorrect Answers

A:

For the technical explanation, that explains why a team might for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable

B:

That explains why a team might conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable.

D:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. General-purpose model duties and downstream application controls are related but not interchangeable.

Question 308

An AI governance reviewer examines this situation: A cloud developer offers a general-purpose text model to multiple downstream application teams. Which acceptance check is most decisive?

  1. Check the legal classification and an obligation-by-obligation evidence matrix before release.
  2. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  3. Compare published promises, disclaimers and actual tested limitations before approval.
  4. Review provider technical documentation and downstream integration requirements under the relevant regime.

Explanation

Correct Answer

D:

For the verification, review provider technical documentation and downstream integration requirements under the relevant regime. This verifies the necessary safeguard rather than merely showing that an AI project was completed. General-purpose model duties and downstream application controls are related but not interchangeable

Incorrect Answers

A:

Check the legal classification and an obligation-by-obligation evidence matrix before release would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

B:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

C:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to review provider technical documentation and downstream integration requirements under the relevant regime.

Question 309

A cloud developer offers a general-purpose text model to multiple downstream application teams. The team instead proposes: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. Which failure should the reviewer expect most directly?

  1. Key information and risk controls may be missing across the model supply chain.
  2. The organization may build a use that cannot lawfully operate in its target context.
  3. Required assurance, documentation or human oversight may be missing at a legal checkpoint.
  4. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.

Explanation

Correct Answer

A:

For the risk comparison, key information and risk controls may be missing across the model supply chain. The shortcut overlooks the central safeguard: identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. General-purpose model duties and downstream application controls are related but not interchangeable

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Attach a generic responsible-AI statement instead of testing specific high-risk requirements. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain

D:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: Key information and risk controls may be missing across the model supply chain.

Question 310

Choose the intervention and validation that fit this business problem: A cloud developer offers a general-purpose text model to multiple downstream application teams.

  1. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  2. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  3. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  4. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

B:

For the action and verification, identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Review provider technical documentation and downstream integration requirements under the relevant regime. Together, they implement and test the intended safeguard. General-purpose model duties and downstream application controls are related but not interchangeable

Incorrect Answers

A:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime.

C:

This pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime.

D:

For the action and verification, this pair supports another governance intervention: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. For the actual case, the required action is: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. The supporting evidence is: Review provider technical documentation and downstream integration requirements under the relevant regime

Question 311

A reseller places a rebranded AI screening product on a regulated market after modifying its workflow. Which step would provide the most targeted control for the problem described?

  1. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  2. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  3. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.
  4. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.

Explanation

Correct Answer

C:

Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market. The practical acceptance evidence is: map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Incorrect Answers

A:

The alternative—identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

B:

The alternative—identify whether an AI interaction or generated content requires specific disclosure in the applicable context—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

D:

The alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 312

The organization chooses to differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law for this situation: A reseller places a rebranded AI screening product on a regulated market after modifying its workflow. Which rationale best supports that choice?

  1. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  2. General-purpose model duties and downstream application controls are related but not interchangeable.
  3. Sensitive data can trigger heightened restrictions and requires particular attention to necessity, safeguards and permitted uses.
  4. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Explanation

Correct Answer

D:

A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market. That is why the case calls for this action: differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.

Incorrect Answers

A:

That explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

B:

That explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

C:

That explains why a team might escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 313

An AI governance reviewer examines this situation: A reseller places a rebranded AI screening product on a regulated market after modifying its workflow. Which acceptance check is most decisive?

  1. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  2. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  3. Check the legal classification and an obligation-by-obligation evidence matrix before release.
  4. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

A:

Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Incorrect Answers

B:

The alternative—simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

C:

The alternative—check the legal classification and an obligation-by-obligation evidence matrix before release—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

D:

The alternative—inspect governance roles, risk tolerance and communication processes that support subsequent risk work—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 314

For this case—A reseller places a rebranded AI screening product on a regulated market after modifying its workflow—staff choose the shortcut: Assume every organization interacting with an AI system is legally only a passive end user. Which consequence best explains why this is inadequate?

  1. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  2. A role change may leave mandatory obligations or documentation unassigned.
  3. People may be misled about interacting with AI or receiving AI-generated material.
  4. Important management or impact-assessment activities may be omitted under the wrong standard.

Explanation

Correct Answer

B:

A role change may leave mandatory obligations or documentation unassigned. The shortcut overlooks the central safeguard: differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned.

C:

This consequence follows more directly from another shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned.

D:

This consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned.

Question 315

The governance team needs both a decision and a way to verify it for this case: A reseller places a rebranded AI screening product on a regulated market after modifying its workflow. Which pair is best?

  1. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  3. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

C:

Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime. Together, they implement and test the intended safeguard. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Incorrect Answers

A:

The alternative—conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

B:

This pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. The supporting evidence is: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

D:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. The supporting evidence is: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Question 316

An AI project team faces this situation: An enterprise deploys an imported supplier model and changes the intended purpose for local use. What should the team do first to manage the central issue?

  1. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.
  2. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.
  3. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.
  4. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.

Explanation

Correct Answer

D:

For the decision, differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market. The practical acceptance evidence is: map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime

Incorrect Answers

A:

The alternative—conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

B:

For the decision, the alternative—identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

C:

The alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 317

A team reviewing the following situation asks for the most relevant governance principle: An enterprise deploys an imported supplier model and changes the intended purpose for local use. Which principle best explains the required control?

  1. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.
  2. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.
  3. General-purpose model duties and downstream application controls are related but not interchangeable.
  4. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.

Explanation

Correct Answer

A:

For the technical explanation, a business's legal obligations can depend on its actual role, including whether it places or operates a system on the market. That is why the case calls for this action: differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law

Incorrect Answers

B:

For the technical explanation, that explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

C:

For the technical explanation, that explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

D:

That explains why a team might review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use, but it does not justify the decision required by the case. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 318

An enterprise deploys an imported supplier model and changes the intended purpose for local use. Which recorded test or observation most directly demonstrates an effective response?

  1. Check the legal classification and an obligation-by-obligation evidence matrix before release.
  2. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  3. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  4. Review license terms and approval records for representative public-facing generated materials.

Explanation

Correct Answer

B:

For the verification, map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

Incorrect Answers

A:

For the verification, the alternative—check the legal classification and an obligation-by-obligation evidence matrix before release—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

C:

The alternative—record the affected people, purpose, jurisdiction, current legal classification and legal disposition—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

D:

The alternative—review license terms and approval records for representative public-facing generated materials—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 319

A team considers this AI use: An enterprise deploys an imported supplier model and changes the intended purpose for local use. Rather than applying the relevant control, it will assume every organization interacting with an AI system is legally only a passive end user. What can go wrong?

  1. The organization may build a use that cannot lawfully operate in its target context.
  2. People may be misled about interacting with AI or receiving AI-generated material.
  3. A role change may leave mandatory obligations or documentation unassigned.
  4. Important management or impact-assessment activities may be omitted under the wrong standard.

Explanation

Correct Answer

C:

For the risk comparison, a role change may leave mandatory obligations or documentation unassigned. The shortcut overlooks the central safeguard: differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Hide the use of a customer-facing synthetic voice because the answer content is plausible. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned

D:

For the risk comparison, this consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: A role change may leave mandatory obligations or documentation unassigned

Question 320

An enterprise deploys an imported supplier model and changes the intended purpose for local use. Which action and evidence pair most directly addresses the situation?

  1. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  3. Action: Include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.
  4. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Explanation

Correct Answer

D:

For the action and verification, differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime. Together, they implement and test the intended safeguard. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

Incorrect Answers

A:

For the action and verification, the alternative—conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market

B:

This pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. The supporting evidence is: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

C:

The alternative—include foreseeable AI failure modes, product interfaces and remedial controls in the relevant safety/legal assessment. Evidence: Review documented hazards, foreseeable misuse, safety testing and corrective action ownership—would not resolve the stated issue. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.

Question 321

Consider the following proposed AI use: A regulated AI provider cannot retrieve the tests used to support its market-release decision. Which response best addresses the governance gap?

  1. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  2. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.
  3. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.
  4. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Explanation

Correct Answer

A:

Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records. The practical acceptance evidence is: simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Incorrect Answers

B:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

C:

The alternative—conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

D:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

Question 322

A regulated AI provider cannot retrieve the tests used to support its market-release decision. Which explanation best connects the risk to the proposed governance action?

  1. General-purpose model duties and downstream application controls are related but not interchangeable.
  2. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  3. A business's legal obligations can depend on its actual role, including whether it places or operates a system on the market.
  4. AI-supported decisions do not automatically remove individuals' legal rights to information, contestation or appropriate review where applicable.

Explanation

Correct Answer

B:

A law may require demonstrable compliance; good intentions are not a substitute for reproducible records. That is why the case calls for this action: keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

Incorrect Answers

A:

That explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

C:

That explains why a team might differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

D:

That explains why a team might design a process to identify relevant data subject requests and assess applicable automated-decision safeguards, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Question 323

The organization has implemented a control for this case: A regulated AI provider cannot retrieve the tests used to support its market-release decision. Which result would best validate it?

  1. Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  2. Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  3. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  4. Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs.

Explanation

Correct Answer

C:

Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Incorrect Answers

A:

Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

B:

Record the affected people, purpose, jurisdiction, current legal classification and legal disposition would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

D:

Test sample requests for access, correction and review against the governing jurisdictional rules and operational logs would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Question 324

A regulated AI provider cannot retrieve the tests used to support its market-release decision. The team instead proposes: Wait for a regulator to ask questions before creating system documentation and records. Which failure should the reviewer expect most directly?

  1. Key information and risk controls may be missing across the model supply chain.
  2. Required assurance, documentation or human oversight may be missing at a legal checkpoint.
  3. Unneeded sensitive fields can leak or influence outcomes while adding no justified benefit.
  4. The organization may be unable to show compliance or respond effectively to a formal inquiry.

Explanation

Correct Answer

D:

The organization may be unable to show compliance or respond effectively to a formal inquiry. The shortcut overlooks the central safeguard: keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry.

B:

This consequence follows more directly from another shortcut: Attach a generic responsible-AI statement instead of testing specific high-risk requirements. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry.

C:

This consequence follows more directly from another shortcut: Feed the entire customer database into the model because storage and compute are available. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry.

Question 325

A regulated AI provider cannot retrieve the tests used to support its market-release decision. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  2. Action: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. Evidence: Test representative user journeys for the required AI notices and correct timing.
  3. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.
  4. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.

Explanation

Correct Answer

A:

Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history. Together, they implement and test the intended safeguard. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Incorrect Answers

B:

This pair supports another governance intervention: Identify whether an AI interaction or generated content requires specific disclosure in the applicable context. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

C:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

D:

This pair supports another governance intervention: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Question 326

An importer receives a compliance request but has no trace of the supplier's technical documentation. Which step would provide the most targeted control for the problem described?

  1. Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development.
  2. Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.
  3. Identify whether an AI interaction or generated content requires specific disclosure in the applicable context.
  4. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Explanation

Correct Answer

B:

For the decision, keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records. The practical acceptance evidence is: simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history

Incorrect Answers

A:

For the decision, the alternative—conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review

C:

The alternative—identify whether an AI interaction or generated content requires specific disclosure in the applicable context—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

D:

The alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—is not the most direct response to this case. The matching action is to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review.

Question 327

The organization chooses to keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review for this situation: An importer receives a compliance request but has no trace of the supplier's technical documentation. Which rationale best supports that choice?

  1. General-purpose model duties and downstream application controls are related but not interchangeable.
  2. A high-risk classification can create obligations beyond a generic fairness pledge, subject to the applicable law and role.
  3. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.
  4. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Explanation

Correct Answer

C:

For the technical explanation, a law may require demonstrable compliance; good intentions are not a substitute for reproducible records. That is why the case calls for this action: keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review

Incorrect Answers

A:

For the technical explanation, that explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records

B:

That explains why a team might for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

D:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Question 328

An importer receives a compliance request but has no trace of the supplier's technical documentation. Which recorded test or observation most directly demonstrates an effective response?

  1. Check the legal classification and an obligation-by-obligation evidence matrix before release.
  2. Review provider technical documentation and downstream integration requirements under the relevant regime.
  3. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  4. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Explanation

Correct Answer

D:

For the verification, simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records

Incorrect Answers

A:

Check the legal classification and an obligation-by-obligation evidence matrix before release would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

B:

Review provider technical documentation and downstream integration requirements under the relevant regime would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

C:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Question 329

An importer receives a compliance request but has no trace of the supplier's technical documentation. The team instead proposes: Wait for a regulator to ask questions before creating system documentation and records. Which failure should the reviewer expect most directly?

  1. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  2. The organization may build a use that cannot lawfully operate in its target context.
  3. Key information and risk controls may be missing across the model supply chain.
  4. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

Explanation

Correct Answer

A:

For the risk comparison, the organization may be unable to show compliance or respond effectively to a formal inquiry. The shortcut overlooks the central safeguard: keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Assume deploying a general-purpose model API creates no obligations for either supplier or integrator. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry

D:

This consequence follows more directly from another shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. The shortcut actually described would instead create this risk: The organization may be unable to show compliance or respond effectively to a formal inquiry.

Question 330

The governance team needs both a decision and a way to verify it for this case: An importer receives a compliance request but has no trace of the supplier's technical documentation. Which pair is best?

  1. Action: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. Evidence: Record the affected people, purpose, jurisdiction, current legal classification and legal disposition.
  2. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.
  3. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.
  4. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

B:

For the action and verification, keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history. Together, they implement and test the intended safeguard. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records

Incorrect Answers

A:

This pair supports another governance intervention: Conduct a jurisdiction- and use-specific prohibited-practice screen before selecting a use case for development. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

C:

This pair supports another governance intervention: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

D:

This pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. The supporting evidence is: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Question 331

A global nonprofit wants a common responsible-AI policy across regions with different local regulations. Which step would provide the most targeted control for the problem described?

  1. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.
  2. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  3. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  4. Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities.

Explanation

Correct Answer

C:

Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification. The practical acceptance evidence is: map internal policy commitments to concrete controls and evidence against the OECD principles.

Incorrect Answers

A:

The alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

B:

The alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

D:

This would help address a separate governance task: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For this case, however, an international principles framework helps establish consistent governance expectations but is not itself a product-specific certification. The more direct step is: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Question 332

A global nonprofit wants a common responsible-AI policy across regions with different local regulations. Which explanation best connects the risk to the proposed governance action?

  1. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  2. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  3. Deceptive descriptions of model capabilities can create legal risk even when the underlying product functions.
  4. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Explanation

Correct Answer

D:

An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification. That is why the case calls for this action: use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Incorrect Answers

A:

That explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

B:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

C:

That explains why a team might review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 333

The organization has implemented a control for this case: A global nonprofit wants a common responsible-AI policy across regions with different local regulations. Which result would best validate it?

  1. Map internal policy commitments to concrete controls and evidence against the OECD principles.
  2. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Review license terms and approval records for representative public-facing generated materials.

Explanation

Correct Answer

A:

Map internal policy commitments to concrete controls and evidence against the OECD principles. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Incorrect Answers

B:

The alternative—inspect governance roles, risk tolerance and communication processes that support subsequent risk work—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

C:

The alternative—trace a selected use case from contextual mapping to measured evidence and a documented risk response—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

D:

The alternative—review license terms and approval records for representative public-facing generated materials—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 334

A global nonprofit wants a common responsible-AI policy across regions with different local regulations. The team instead proposes: Treat the existence of an OECD policy reference as proof that a particular model has passed technical assurance. Which failure should the reviewer expect most directly?

  1. Important management or impact-assessment activities may be omitted under the wrong standard.
  2. High-level aspirations may not translate into tested safeguards or accountability.
  3. Technical scores do not create decision ownership or a risk governance structure.
  4. Public outputs can create contractual or infringement exposure even without intentional copying.

Explanation

Correct Answer

B:

High-level aspirations may not translate into tested safeguards or accountability. The shortcut overlooks the central safeguard: use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability.

C:

This consequence follows more directly from another shortcut: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability.

D:

This consequence follows more directly from another shortcut: Assume every generated image is automatically free of third-party rights or usage limitations. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability.

Question 335

A global nonprofit wants a common responsible-AI policy across regions with different local regulations. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  2. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  3. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  4. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Explanation

Correct Answer

C:

Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Map internal policy commitments to concrete controls and evidence against the OECD principles. Together, they implement and test the intended safeguard. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Incorrect Answers

A:

The alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

B:

The alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

D:

The alternative—keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 336

An AI project team faces this situation: An international shipping group needs shared AI principles before tailoring controls to its operations. What should the team do first to manage the central issue?

  1. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.
  2. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.
  3. Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law.
  4. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.

Explanation

Correct Answer

D:

For the decision, use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification. The practical acceptance evidence is: map internal policy commitments to concrete controls and evidence against the OECD principles

Incorrect Answers

A:

For the decision, the alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

B:

The alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

C:

The alternative—differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 337

The organization chooses to use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability for this situation: An international shipping group needs shared AI principles before tailoring controls to its operations. Which rationale best supports that choice?

  1. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.
  2. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  3. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  4. General-purpose model duties and downstream application controls are related but not interchangeable.

Explanation

Correct Answer

A:

For the technical explanation, an international principles framework helps establish consistent governance expectations but is not itself a product-specific certification. That is why the case calls for this action: use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability

Incorrect Answers

B:

That explains why a team might establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

C:

For the technical explanation, that explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

D:

That explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 338

An international shipping group needs shared AI principles before tailoring controls to its operations. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Map internal policy commitments to concrete controls and evidence against the OECD principles.
  3. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Review documented hazards, foreseeable misuse, safety testing and corrective action ownership.

Explanation

Correct Answer

B:

For the verification, map internal policy commitments to concrete controls and evidence against the OECD principles. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

Incorrect Answers

A:

The alternative—verify that the chosen reference matches the control objective and that the project cites the correct standard—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

C:

For the verification, the alternative—trace a selected use case from contextual mapping to measured evidence and a documented risk response—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

D:

The alternative—review documented hazards, foreseeable misuse, safety testing and corrective action ownership—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

Question 339

For this case—An international shipping group needs shared AI principles before tailoring controls to its operations—staff choose the shortcut: Treat the existence of an OECD policy reference as proof that a particular model has passed technical assurance. Which consequence best explains why this is inadequate?

  1. The organization can overlook use context and fail to act on significant measured risks.
  2. Important management or impact-assessment activities may be omitted under the wrong standard.
  3. High-level aspirations may not translate into tested safeguards or accountability.
  4. People's data may support decisions they were never reasonably told to expect.

Explanation

Correct Answer

C:

For the risk comparison, high-level aspirations may not translate into tested safeguards or accountability. The shortcut overlooks the central safeguard: use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability

D:

This consequence follows more directly from another shortcut: Declare every internal AI experiment a permitted extension of the original collection purpose. The shortcut actually described would instead create this risk: High-level aspirations may not translate into tested safeguards or accountability.

Question 340

An international shipping group needs shared AI principles before tailoring controls to its operations. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Action: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. Evidence: Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  4. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.

Explanation

Correct Answer

D:

For the action and verification, use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Map internal policy commitments to concrete controls and evidence against the OECD principles. Together, they implement and test the intended safeguard. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

Incorrect Answers

A:

The alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.

B:

For the action and verification, the alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work—would not resolve the stated issue. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification

C:

This pair supports another governance intervention: Review hiring-model outcomes for unjustified disparate treatment and applicable nondiscrimination requirements. For the actual case, the required action is: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. The supporting evidence is: Map internal policy commitments to concrete controls and evidence against the OECD principles.

Question 341

A public laboratory adopts NIST AI RMF but has no assigned authority for risk acceptance. Which governance action most directly addresses the need in this situation?

  1. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.
  2. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  3. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  4. Document rights and restrictions for copyrighted or licensed sources before training or adapting a model.

Explanation

Correct Answer

A:

Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization. The practical acceptance evidence is: inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Incorrect Answers

B:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

C:

The alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

D:

The alternative—document rights and restrictions for copyrighted or licensed sources before training or adapting a model—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Question 342

A public laboratory adopts NIST AI RMF but has no assigned authority for risk acceptance. Which explanation best connects the risk to the proposed governance action?

  1. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  2. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  3. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  4. Transparency duties depend on the system, audience and jurisdiction, not a universal requirement to publish model weights.

Explanation

Correct Answer

B:

GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization. That is why the case calls for this action: establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.

Incorrect Answers

A:

That explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

C:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

D:

That explains why a team might identify whether an AI interaction or generated content requires specific disclosure in the applicable context, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Question 343

The organization has implemented a control for this case: A public laboratory adopts NIST AI RMF but has no assigned authority for risk acceptance. Which result would best validate it?

  1. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  3. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  4. Compare published promises, disclaimers and actual tested limitations before approval.

Explanation

Correct Answer

C:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work. This verifies the necessary safeguard rather than merely showing that an AI project was completed. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Incorrect Answers

A:

Trace a selected use case from contextual mapping to measured evidence and a documented risk response would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

B:

Verify that the chosen reference matches the control objective and that the project cites the correct standard would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

D:

Compare published promises, disclaimers and actual tested limitations before approval would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Question 344

A public laboratory adopts NIST AI RMF but has no assigned authority for risk acceptance. The team instead proposes: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. Which failure should the reviewer expect most directly?

  1. The organization can overlook use context and fail to act on significant measured risks.
  2. Important management or impact-assessment activities may be omitted under the wrong standard.
  3. A role change may leave mandatory obligations or documentation unassigned.
  4. Technical scores do not create decision ownership or a risk governance structure.

Explanation

Correct Answer

D:

Technical scores do not create decision ownership or a risk governance structure. The shortcut overlooks the central safeguard: establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure.

B:

This consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure.

C:

This consequence follows more directly from another shortcut: Assume every organization interacting with an AI system is legally only a passive end user. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure.

Question 345

Choose the intervention and validation that fit this business problem: A public laboratory adopts NIST AI RMF but has no assigned authority for risk acceptance.

  1. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  2. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  3. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Action: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. Evidence: Check the legal classification and an obligation-by-obligation evidence matrix before release.

Explanation

Correct Answer

A:

Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work. Together, they implement and test the intended safeguard. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Incorrect Answers

B:

This pair supports another governance intervention: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

C:

This pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

D:

This pair supports another governance intervention: For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Question 346

Consider the following proposed AI use: A retailer maps model metrics but cannot identify who approves exceptions to its AI policy. Which response best addresses the governance gap?

  1. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  2. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.
  3. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  4. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Explanation

Correct Answer

B:

For the decision, establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization. The practical acceptance evidence is: inspect governance roles, risk tolerance and communication processes that support subsequent risk work

Incorrect Answers

A:

For the decision, the alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

C:

For the decision, the alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

D:

The alternative—check the provider's contract and relevant law for permitted commercial use and responsibility for generated content—would not resolve the stated issue. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Question 347

A retailer maps model metrics but cannot identify who approves exceptions to its AI policy. Which explanation best connects the risk to the proposed governance action?

  1. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  2. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.
  3. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  4. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Explanation

Correct Answer

C:

For the technical explanation, gOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization. That is why the case calls for this action: establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function

Incorrect Answers

A:

For the technical explanation, that explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

B:

That explains why a team might use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

D:

That explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.

Question 348

An AI governance reviewer examines this situation: A retailer maps model metrics but cannot identify who approves exceptions to its AI policy. Which acceptance check is most decisive?

  1. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Map internal policy commitments to concrete controls and evidence against the OECD principles.
  3. Evaluate selection and error rates across legally relevant groups, and document response to disparities.
  4. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Explanation

Correct Answer

D:

For the verification, inspect governance roles, risk tolerance and communication processes that support subsequent risk work. This verifies the necessary safeguard rather than merely showing that an AI project was completed. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

Incorrect Answers

A:

For the verification, verify that the chosen reference matches the control objective and that the project cites the correct standard would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work

B:

Map internal policy commitments to concrete controls and evidence against the OECD principles would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

C:

Evaluate selection and error rates across legally relevant groups, and document response to disparities would test a different outcome. The directly relevant evidence is to inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Question 349

A retailer maps model metrics but cannot identify who approves exceptions to its AI policy. The team instead proposes: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. Which failure should the reviewer expect most directly?

  1. Technical scores do not create decision ownership or a risk governance structure.
  2. Important management or impact-assessment activities may be omitted under the wrong standard.
  3. The organization can overlook use context and fail to act on significant measured risks.
  4. Customers may make costly decisions based on unsupported performance claims.

Explanation

Correct Answer

A:

For the risk comparison, technical scores do not create decision ownership or a risk governance structure. The shortcut overlooks the central safeguard: establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure

C:

For the risk comparison, this consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure

D:

This consequence follows more directly from another shortcut: Advertise a probabilistic assistant as error-free because a demo produced a correct answer. The shortcut actually described would instead create this risk: Technical scores do not create decision ownership or a risk governance structure.

Question 350

A retailer maps model metrics but cannot identify who approves exceptions to its AI policy. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  2. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Action: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. Evidence: Simulate an evidence request and locate current classifications, test reports, accountable owners and remediation history.

Explanation

Correct Answer

B:

For the action and verification, establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work. Together, they implement and test the intended safeguard. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work

C:

For the action and verification, this pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work

D:

This pair supports another governance intervention: Keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review. For the actual case, the required action is: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. The supporting evidence is: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.

Question 351

A credit union documents model accuracy but has not mapped who might be harmed or how to respond. Which governance action most directly addresses the need in this situation?

  1. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  2. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  3. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.
  4. Test whether a proposed new model purpose is compatible with the reason personal information was originally obtained.

Explanation

Correct Answer

C:

Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings. The practical acceptance evidence is: trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Incorrect Answers

A:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

B:

The alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

D:

The alternative—test whether a proposed new model purpose is compatible with the reason personal information was originally obtained—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

Question 352

A credit union documents model accuracy but has not mapped who might be harmed or how to respond. Which explanation best connects the risk to the proposed governance action?

  1. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.
  2. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  3. Having technical access to records does not, by itself, authorize a new AI training purpose.
  4. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Explanation

Correct Answer

D:

These functions provide complementary steps: establish context, assess relevant risks and respond to the findings. That is why the case calls for this action: map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

Incorrect Answers

A:

That explains why a team might use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

B:

That explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

C:

That explains why a team might identify and document the applicable lawful basis before reusing personal information for model development, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Question 353

A credit union documents model accuracy but has not mapped who might be harmed or how to respond. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  4. Record the original notice and collection purpose alongside the proposed use and compatibility determination.

Explanation

Correct Answer

A:

Trace a selected use case from contextual mapping to measured evidence and a documented risk response. This verifies the necessary safeguard rather than merely showing that an AI project was completed. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Incorrect Answers

B:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response.

C:

Verify that the chosen reference matches the control objective and that the project cites the correct standard would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response.

D:

Record the original notice and collection purpose alongside the proposed use and compatibility determination would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Question 354

A credit union documents model accuracy but has not mapped who might be harmed or how to respond. The team instead proposes: Call a single model performance benchmark a complete application of every risk-management function. Which failure should the reviewer expect most directly?

  1. High-level aspirations may not translate into tested safeguards or accountability.
  2. The organization can overlook use context and fail to act on significant measured risks.
  3. Important management or impact-assessment activities may be omitted under the wrong standard.
  4. Identifiable or sensitive characteristics may still be exposed or processed unlawfully.

Explanation

Correct Answer

B:

The organization can overlook use context and fail to act on significant measured risks. The shortcut overlooks the central safeguard: map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat the existence of an OECD policy reference as proof that a particular model has passed technical assurance. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks.

C:

This consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks.

D:

This consequence follows more directly from another shortcut: Treat face embeddings as ordinary anonymous telemetry simply because names are removed. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks.

Question 355

A credit union documents model accuracy but has not mapped who might be harmed or how to respond. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  4. Action: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. Evidence: Confirm the applicable category, explicit necessity assessment, protective controls and approved processing conditions.

Explanation

Correct Answer

C:

Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Trace a selected use case from contextual mapping to measured evidence and a documented risk response. Together, they implement and test the intended safeguard. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Incorrect Answers

A:

This pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

B:

This pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

D:

This pair supports another governance intervention: Escalate biometric or other specially protected personal data for a jurisdiction-specific legal and security review. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Question 356

Consider the following proposed AI use: A hospital tracks an AI triage model's latency without defining the operational risks and mitigations. Which response best addresses the governance gap?

  1. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  2. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  3. Remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient.
  4. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

Explanation

Correct Answer

D:

For the decision, map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings. The practical acceptance evidence is: trace a selected use case from contextual mapping to measured evidence and a documented risk response

Incorrect Answers

A:

For the decision, the alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions

B:

For the decision, the alternative—select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions

C:

The alternative—remove unnecessary identifiers and features before training and inference, while testing whether utility remains sufficient—is not the most direct response to this case. The matching action is to map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.

Question 357

A hospital tracks an AI triage model's latency without defining the operational risks and mitigations. Why is the recommended governance approach appropriate?

  1. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  2. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  3. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  4. General-purpose model duties and downstream application controls are related but not interchangeable.

Explanation

Correct Answer

A:

For the technical explanation, these functions provide complementary steps: establish context, assess relevant risks and respond to the findings. That is why the case calls for this action: map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions

Incorrect Answers

B:

For the technical explanation, that explains why a team might select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings

C:

That explains why a team might establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

D:

That explains why a team might identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities, but it does not justify the decision required by the case. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.

Question 358

A hospital tracks an AI triage model's latency without defining the operational risks and mitigations. Which recorded test or observation most directly demonstrates an effective response?

  1. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  3. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  4. Test representative user journeys for the required AI notices and correct timing.

Explanation

Correct Answer

B:

For the verification, trace a selected use case from contextual mapping to measured evidence and a documented risk response. This verifies the necessary safeguard rather than merely showing that an AI project was completed. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings

Incorrect Answers

A:

For the verification, verify that the chosen reference matches the control objective and that the project cites the correct standard would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response

C:

For the verification, inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response

D:

Test representative user journeys for the required AI notices and correct timing would test a different outcome. The directly relevant evidence is to trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Question 359

The organization has this problem: A hospital tracks an AI triage model's latency without defining the operational risks and mitigations. It takes a shortcut: Call a single model performance benchmark a complete application of every risk-management function. What is the principal risk?

  1. Technical scores do not create decision ownership or a risk governance structure.
  2. Important management or impact-assessment activities may be omitted under the wrong standard.
  3. The organization can overlook use context and fail to act on significant measured risks.
  4. Proxy effects can cause systematically unequal opportunities even with no protected attribute field.

Explanation

Correct Answer

C:

For the risk comparison, the organization can overlook use context and fail to act on significant measured risks. The shortcut overlooks the central safeguard: map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks

D:

This consequence follows more directly from another shortcut: Use overall model accuracy alone as proof that a recruitment tool is legally fair. The shortcut actually described would instead create this risk: The organization can overlook use context and fail to act on significant measured risks.

Question 360

The governance team needs both a decision and a way to verify it for this case: A hospital tracks an AI triage model's latency without defining the operational risks and mitigations. Which pair is best?

  1. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Action: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. Evidence: Review provider technical documentation and downstream integration requirements under the relevant regime.
  4. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Explanation

Correct Answer

D:

For the action and verification, map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Trace a selected use case from contextual mapping to measured evidence and a documented risk response. Together, they implement and test the intended safeguard. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response

B:

For the action and verification, this pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response

C:

This pair supports another governance intervention: Identify the model provider's and downstream integrator's applicable general-purpose AI responsibilities. For the actual case, the required action is: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. The supporting evidence is: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.

Question 361

Consider the following proposed AI use: A financial group selects an ISO reference for its organization-wide AI governance management system. Which response best addresses the governance gap?

  1. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  2. Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function.
  3. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.
  4. Check the provider's contract and relevant law for permitted commercial use and responsibility for generated content.

Explanation

Correct Answer

A:

Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement. The practical acceptance evidence is: verify that the chosen reference matches the control objective and that the project cites the correct standard.

Incorrect Answers

B:

The alternative—establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

C:

The alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

D:

The alternative—check the provider's contract and relevant law for permitted commercial use and responsibility for generated content—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

Question 362

A financial group selects an ISO reference for its organization-wide AI governance management system. Why is the recommended governance approach appropriate?

  1. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  2. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  3. An international principles framework helps establish consistent governance expectations but is not itself a product-specific certification.
  4. Public visibility of material does not establish blanket rights to copy or use it for every machine-learning purpose.

Explanation

Correct Answer

B:

Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement. That is why the case calls for this action: select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

Incorrect Answers

A:

That explains why a team might establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

C:

That explains why a team might use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

D:

That explains why a team might document rights and restrictions for copyrighted or licensed sources before training or adapting a model, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Question 363

A financial group selects an ISO reference for its organization-wide AI governance management system. Which recorded test or observation most directly demonstrates an effective response?

  1. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Map internal policy commitments to concrete controls and evidence against the OECD principles.
  3. Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  4. Test representative user journeys for the required AI notices and correct timing.

Explanation

Correct Answer

C:

Verify that the chosen reference matches the control objective and that the project cites the correct standard. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Incorrect Answers

A:

Trace a selected use case from contextual mapping to measured evidence and a documented risk response would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard.

B:

Map internal policy commitments to concrete controls and evidence against the OECD principles would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard.

D:

Test representative user journeys for the required AI notices and correct timing would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard.

Question 364

The organization has this problem: A financial group selects an ISO reference for its organization-wide AI governance management system. It takes a shortcut: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. What is the principal risk?

  1. High-level aspirations may not translate into tested safeguards or accountability.
  2. The organization can overlook use context and fail to act on significant measured risks.
  3. The organization may be unable to show compliance or respond effectively to a formal inquiry.
  4. Important management or impact-assessment activities may be omitted under the wrong standard.

Explanation

Correct Answer

D:

Important management or impact-assessment activities may be omitted under the wrong standard. The shortcut overlooks the central safeguard: select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat the existence of an OECD policy reference as proof that a particular model has passed technical assurance. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard.

B:

This consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard.

C:

This consequence follows more directly from another shortcut: Wait for a regulator to ask questions before creating system documentation and records. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard.

Question 365

Choose the intervention and validation that fit this business problem: A financial group selects an ISO reference for its organization-wide AI governance management system.

  1. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  2. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  3. Action: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. Evidence: Map internal policy commitments to concrete controls and evidence against the OECD principles.
  4. Action: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. Evidence: Compare published promises, disclaimers and actual tested limitations before approval.

Explanation

Correct Answer

A:

Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Verify that the chosen reference matches the control objective and that the project cites the correct standard. Together, they implement and test the intended safeguard. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Incorrect Answers

B:

This pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

C:

This pair supports another governance intervention: Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

D:

This pair supports another governance intervention: Review AI-related consumer claims for accuracy, substantiation and misleading omissions before marketing or use. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Question 366

An AI project team faces this situation: A city agency needs a standard to structure AI-system impact assessment without confusing it with terminology guidance. What should the team do first to manage the central issue?

  1. Use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability.
  2. Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.
  3. Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions.
  4. For an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls.

Explanation

Correct Answer

B:

For the decision, select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement. The practical acceptance evidence is: verify that the chosen reference matches the control objective and that the project cites the correct standard

Incorrect Answers

A:

The alternative—use the OECD trustworthy-AI principles as a policy-level reference for human-centered values, transparency, robustness and accountability—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

C:

For the decision, the alternative—map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance

D:

The alternative—for an AI use classified high-risk under an applicable regime, plan required risk, data, documentation and oversight controls—is not the most direct response to this case. The matching action is to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance.

Question 367

The organization chooses to select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance for this situation: A city agency needs a standard to structure AI-system impact assessment without confusing it with terminology guidance. Which rationale best supports that choice?

  1. These functions provide complementary steps: establish context, assess relevant risks and respond to the findings.
  2. GOVERN provides cross-cutting structures that guide how AI risk work is carried out across the organization.
  3. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.
  4. A law may require demonstrable compliance; good intentions are not a substitute for reproducible records.

Explanation

Correct Answer

C:

For the technical explanation, different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement. That is why the case calls for this action: select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance

Incorrect Answers

A:

That explains why a team might map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

B:

For the technical explanation, that explains why a team might establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement

D:

That explains why a team might keep a current evidence map of relevant requirements, assessments and decision records for regulator or conformity review, but it does not justify the decision required by the case. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement.

Question 368

An AI governance reviewer examines this situation: A city agency needs a standard to structure AI-system impact assessment without confusing it with terminology guidance. Which acceptance check is most decisive?

  1. Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  3. Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion.
  4. Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Explanation

Correct Answer

D:

For the verification, verify that the chosen reference matches the control objective and that the project cites the correct standard. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement

Incorrect Answers

A:

For the verification, trace a selected use case from contextual mapping to measured evidence and a documented risk response would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard

B:

Inspect governance roles, risk tolerance and communication processes that support subsequent risk work would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard.

C:

Check the data-use assessment for jurisdiction, purpose, applicable lawful basis and approval before ingestion would test a different outcome. The directly relevant evidence is to verify that the chosen reference matches the control objective and that the project cites the correct standard.

Question 369

A city agency needs a standard to structure AI-system impact assessment without confusing it with terminology guidance. The team instead proposes: Claim ISO/IEC 22989 alone establishes an audited AI management system or completes an impact assessment. Which failure should the reviewer expect most directly?

  1. Important management or impact-assessment activities may be omitted under the wrong standard.
  2. The organization can overlook use context and fail to act on significant measured risks.
  3. Technical scores do not create decision ownership or a risk governance structure.
  4. The organization may build a use that cannot lawfully operate in its target context.

Explanation

Correct Answer

A:

For the risk comparison, important management or impact-assessment activities may be omitted under the wrong standard. The shortcut overlooks the central safeguard: select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Call a single model performance benchmark a complete application of every risk-management function. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard

C:

This consequence follows more directly from another shortcut: Replace organizational accountability with a model accuracy chart and call it the GOVERN function. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard.

D:

This consequence follows more directly from another shortcut: Proceed with a sensitive behavior-profiling use case first and check prohibitions only after launch. The shortcut actually described would instead create this risk: Important management or impact-assessment activities may be omitted under the wrong standard.

Question 370

A city agency needs a standard to structure AI-system impact assessment without confusing it with terminology guidance. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. Evidence: Trace a selected use case from contextual mapping to measured evidence and a documented risk response.
  2. Action: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Evidence: Verify that the chosen reference matches the control objective and that the project cites the correct standard.
  3. Action: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. Evidence: Inspect governance roles, risk tolerance and communication processes that support subsequent risk work.
  4. Action: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. Evidence: Map contracts, modifications, branding and deployment acts to the role definitions used in the relevant regime.

Explanation

Correct Answer

B:

For the action and verification, select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. Verify that the chosen reference matches the control objective and that the project cites the correct standard. Together, they implement and test the intended safeguard. Different ISO documents serve different purposes; a vocabulary standard is not the same as a management-system requirement

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Map AI context, measure identified risks and manage prioritized responses using the NIST AI RMF functions. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard

C:

This pair supports another governance intervention: Establish the organizational policies, roles and oversight practices associated with the NIST AI RMF GOVERN function. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

D:

This pair supports another governance intervention: Differentiate provider, deployer, importer and distributor responsibilities for the AI supply arrangement under applicable law. For the actual case, the required action is: Select ISO/IEC 42001 for AI management-system governance, ISO/IEC 22989 for terminology and concepts, and ISO/IEC 42005 for AI-system impact assessment guidance. The supporting evidence is: Verify that the chosen reference matches the control objective and that the project cites the correct standard.

Question 371

A municipal council wants an AI translator but has not decided which official communications it may handle. Which governance action most directly addresses the need in this situation?

  1. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  2. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  3. Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.
  4. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

Explanation

Correct Answer

C:

Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. The practical acceptance evidence is: validate the use-case specification with business and affected-user representatives before model selection.

Incorrect Answers

A:

This would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. The more direct step is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

B:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. The more direct step is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

D:

The alternative—maintain accurate user instructions, technical limitations and any required public disclosures for the released model—would not resolve the stated issue. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Question 372

A municipal council wants an AI translator but has not decided which official communications it may handle. Why is the recommended governance approach appropriate?

  1. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  2. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  3. Real-world distributions can change after launch, making old validation evidence insufficient.
  4. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Explanation

Correct Answer

D:

Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. That is why the case calls for this action: write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

Incorrect Answers

A:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

B:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

C:

That explains why a team might set post-release alerts for changes in input populations, outcomes and model quality, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Question 373

The organization has implemented a control for this case: A municipal council wants an AI translator but has not decided which official communications it may handle. Which result would best validate it?

  1. Validate the use-case specification with business and affected-user representatives before model selection.
  2. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  3. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

A:

Validate the use-case specification with business and affected-user representatives before model selection. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Incorrect Answers

B:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

C:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

D:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

Question 374

A team considers this AI use: A municipal council wants an AI translator but has not decided which official communications it may handle. Rather than applying the relevant control, it will choose the highest-scoring model benchmark first and define the business objective afterward. What can go wrong?

  1. Auditors and incident responders cannot recover why critical choices were made.
  2. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  3. Low-frequency, high-impact harms can be neglected.
  4. Subsequent drift or control failures can remain unnoticed.

Explanation

Correct Answer

B:

The solution may optimize an irrelevant metric while exposing people to unassessed harms. The shortcut overlooks the central safeguard: write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

C:

This consequence follows more directly from another shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

D:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

Question 375

Choose the intervention and validation that fit this business problem: A municipal council wants an AI translator but has not decided which official communications it may handle.

  1. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  3. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.
  4. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Explanation

Correct Answer

C:

Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Validate the use-case specification with business and affected-user representatives before model selection. Together, they implement and test the intended safeguard. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Incorrect Answers

A:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection.

B:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection.

D:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection.

Question 376

Consider the following proposed AI use: A logistics firm buys a route optimizer without specifying whether safety or delivery speed is the main goal. Which response best addresses the governance gap?

  1. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  2. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.
  3. Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.
  4. Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

Explanation

Correct Answer

D:

For the decision, write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. The practical acceptance evidence is: validate the use-case specification with business and affected-user representatives before model selection

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. The more direct step is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria

B:

The alternative—evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open—would not resolve the stated issue. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

C:

The alternative—run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate—would not resolve the stated issue. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Question 377

The organization chooses to write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria for this situation: A logistics firm buys a route optimizer without specifying whether safety or delivery speed is the main goal. Which rationale best supports that choice?

  1. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.
  2. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  3. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  4. A dataset can be large yet systematically miss important subgroups or operating conditions.

Explanation

Correct Answer

A:

For the technical explanation, governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence. That is why the case calls for this action: write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria

Incorrect Answers

B:

For the technical explanation, that explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence

C:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

D:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Question 378

A logistics firm buys a route optimizer without specifying whether safety or delivery speed is the main goal. Which recorded test or observation most directly demonstrates an effective response?

  1. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Validate the use-case specification with business and affected-user representatives before model selection.
  3. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Sample labeled records across sources and measure disagreement, missingness and remediation results.

Explanation

Correct Answer

B:

For the verification, validate the use-case specification with business and affected-user representatives before model selection. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence

Incorrect Answers

A:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

C:

Review a documented impact assessment with risk owners, affected groups and recorded design responses would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

D:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to validate the use-case specification with business and affected-user representatives before model selection.

Question 379

The organization has this problem: A logistics firm buys a route optimizer without specifying whether safety or delivery speed is the main goal. It takes a shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. What is the principal risk?

  1. Human involvement becomes a rubber stamp rather than an effective safeguard.
  2. The design can ignore burdens imposed on people who never chose the tool.
  3. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  4. Prompt abuse, unsafe escalation and policy bypass may remain undetected.

Explanation

Correct Answer

C:

For the risk comparison, the solution may optimize an irrelevant metric while exposing people to unassessed harms. The shortcut overlooks the central safeguard: write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Add a nominal approval button while requiring staff to accept results within seconds. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

B:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

D:

This consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: The solution may optimize an irrelevant metric while exposing people to unassessed harms.

Question 380

Choose the intervention and validation that fit this business problem: A logistics firm buys a route optimizer without specifying whether safety or delivery speed is the main goal.

  1. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  3. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  4. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.

Explanation

Correct Answer

D:

For the action and verification, write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Validate the use-case specification with business and affected-user representatives before model selection. Together, they implement and test the intended safeguard. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection

B:

This pair supports another governance intervention: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection.

C:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. The supporting evidence is: Validate the use-case specification with business and affected-user representatives before model selection.

Question 381

A housing authority designs a waitlist prioritizer with input only from contract managers. Which step would provide the most targeted control for the problem described?

  1. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  2. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  3. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  4. Require notification, evaluation and approval paths for significant updates to externally supplied models.

Explanation

Correct Answer

A:

Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The system's users, decision subjects and impacted bystanders can experience different benefits and harms. The practical acceptance evidence is: check stakeholder maps and feedback records for representation of affected groups and mitigations.

Incorrect Answers

B:

This would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

C:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

D:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

Question 382

A housing authority designs a waitlist prioritizer with input only from contract managers. Which explanation best connects the risk to the proposed governance action?

  1. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  2. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.
  3. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.
  4. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Explanation

Correct Answer

B:

The system's users, decision subjects and impacted bystanders can experience different benefits and harms. That is why the case calls for this action: identify directly and indirectly affected groups and incorporate their concerns into design reviews.

Incorrect Answers

A:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

C:

That explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

D:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Question 383

An AI governance reviewer examines this situation: A housing authority designs a waitlist prioritizer with input only from contract managers. Which acceptance check is most decisive?

  1. Validate the use-case specification with business and affected-user representatives before model selection.
  2. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  3. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  4. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Explanation

Correct Answer

C:

Check stakeholder maps and feedback records for representation of affected groups and mitigations. This verifies the necessary safeguard rather than merely showing that an AI project was completed. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Incorrect Answers

A:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

B:

Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

D:

Trace selected training records back to licensed or approved origins through recorded preprocessing steps would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

Question 384

A housing authority designs a waitlist prioritizer with input only from contract managers. The team instead proposes: Consult only the purchaser because it funds the AI deployment. Which failure should the reviewer expect most directly?

  1. Human involvement becomes a rubber stamp rather than an effective safeguard.
  2. Known risks may remain uncontrolled despite formal notice to users.
  3. Operators cannot reliably reverse a harmful release or identify what changed.
  4. The design can ignore burdens imposed on people who never chose the tool.

Explanation

Correct Answer

D:

The design can ignore burdens imposed on people who never chose the tool. The shortcut overlooks the central safeguard: identify directly and indirectly affected groups and incorporate their concerns into design reviews. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Add a nominal approval button while requiring staff to accept results within seconds. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool.

B:

This consequence follows more directly from another shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool.

C:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool.

Question 385

The governance team needs both a decision and a way to verify it for this case: A housing authority designs a waitlist prioritizer with input only from contract managers. Which pair is best?

  1. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Action: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Evidence: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.
  4. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

A:

Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Check stakeholder maps and feedback records for representation of affected groups and mitigations. Together, they implement and test the intended safeguard. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Incorrect Answers

B:

This pair supports another governance intervention: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

C:

This pair supports another governance intervention: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

D:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Question 386

Consider the following proposed AI use: A transit agency tests automated crowd alerts without speaking to passengers with accessibility needs. Which response best addresses the governance gap?

  1. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  2. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  3. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.
  4. Version models, prompts and data transformations and define tested rollback criteria.

Explanation

Correct Answer

B:

For the decision, identify directly and indirectly affected groups and incorporate their concerns into design reviews. The system's users, decision subjects and impacted bystanders can experience different benefits and harms. The practical acceptance evidence is: check stakeholder maps and feedback records for representation of affected groups and mitigations

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews

C:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

D:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. The more direct step is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

Question 387

The organization chooses to identify directly and indirectly affected groups and incorporate their concerns into design reviews for this situation: A transit agency tests automated crowd alerts without speaking to passengers with accessibility needs. Which rationale best supports that choice?

  1. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  2. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.
  3. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.
  4. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Explanation

Correct Answer

C:

For the technical explanation, the system's users, decision subjects and impacted bystanders can experience different benefits and harms. That is why the case calls for this action: identify directly and indirectly affected groups and incorporate their concerns into design reviews

Incorrect Answers

A:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

B:

For the technical explanation, that explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms

D:

That explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Question 388

A transit agency tests automated crowd alerts without speaking to passengers with accessibility needs. Which recorded test or observation most directly demonstrates an effective response?

  1. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  2. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  3. Compare group and context distributions against intended deployment, then test known underrepresented cases.
  4. Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

D:

For the verification, check stakeholder maps and feedback records for representation of affected groups and mitigations. This verifies the necessary safeguard rather than merely showing that an AI project was completed. The system's users, decision subjects and impacted bystanders can experience different benefits and harms

Incorrect Answers

A:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

B:

Review a documented impact assessment with risk owners, affected groups and recorded design responses would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

C:

Compare group and context distributions against intended deployment, then test known underrepresented cases would test a different outcome. The directly relevant evidence is to check stakeholder maps and feedback records for representation of affected groups and mitigations.

Question 389

A transit agency tests automated crowd alerts without speaking to passengers with accessibility needs. The team instead proposes: Consult only the purchaser because it funds the AI deployment. Which failure should the reviewer expect most directly?

  1. The design can ignore burdens imposed on people who never chose the tool.
  2. Known risks may remain uncontrolled despite formal notice to users.
  3. Auditors and incident responders cannot recover why critical choices were made.
  4. Subsequent drift or control failures can remain unnoticed.

Explanation

Correct Answer

A:

For the risk comparison, the design can ignore burdens imposed on people who never chose the tool. The shortcut overlooks the central safeguard: identify directly and indirectly affected groups and incorporate their concerns into design reviews. The system's users, decision subjects and impacted bystanders can experience different benefits and harms

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool

C:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool.

D:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: The design can ignore burdens imposed on people who never chose the tool.

Question 390

A transit agency tests automated crowd alerts without speaking to passengers with accessibility needs. Which action and evidence pair most directly addresses the situation?

  1. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  2. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  3. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  4. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.

Explanation

Correct Answer

B:

For the action and verification, identify directly and indirectly affected groups and incorporate their concerns into design reviews. Check stakeholder maps and feedback records for representation of affected groups and mitigations. Together, they implement and test the intended safeguard. The system's users, decision subjects and impacted bystanders can experience different benefits and harms

Incorrect Answers

A:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

C:

This pair supports another governance intervention: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

D:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. The supporting evidence is: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Question 391

A student-support model uses attendance history to flag risk without first examining stigma or false positives. Which step would provide the most targeted control for the problem described?

  1. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  2. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.
  3. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.
  4. Prevent leakage of future information or duplicate records across training and independent evaluation sets.

Explanation

Correct Answer

C:

Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. The practical acceptance evidence is: review a documented impact assessment with risk owners, affected groups and recorded design responses.

Incorrect Answers

A:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. The more direct step is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

B:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. The more direct step is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

D:

The alternative—prevent leakage of future information or duplicate records across training and independent evaluation sets—would not resolve the stated issue. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

Question 392

The organization chooses to evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open for this situation: A student-support model uses attendance history to flag risk without first examining stigma or false positives. Which rationale best supports that choice?

  1. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.
  2. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  3. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.
  4. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Explanation

Correct Answer

D:

Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. That is why the case calls for this action: evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

Incorrect Answers

A:

That explains why a team might write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

B:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

C:

That explains why a team might run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Question 393

An AI governance reviewer examines this situation: A student-support model uses attendance history to flag risk without first examining stigma or false positives. Which acceptance check is most decisive?

  1. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  2. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  3. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  4. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Explanation

Correct Answer

A:

Review a documented impact assessment with risk owners, affected groups and recorded design responses. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Incorrect Answers

B:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

C:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

D:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

Question 394

A team considers this AI use: A student-support model uses attendance history to flag risk without first examining stigma or false positives. Rather than applying the relevant control, it will wait to assess foreseeable harms until an AI system has completed its first production cycle. What can go wrong?

  1. Known risks may remain uncontrolled despite formal notice to users.
  2. Serious issues can become embedded in model features and operating procedures.
  3. Human involvement becomes a rubber stamp rather than an effective safeguard.
  4. Operators cannot reliably reverse a harmful release or identify what changed.

Explanation

Correct Answer

B:

Serious issues can become embedded in model features and operating procedures. The shortcut overlooks the central safeguard: evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

C:

This consequence follows more directly from another shortcut: Add a nominal approval button while requiring staff to accept results within seconds. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

D:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

Question 395

Choose the intervention and validation that fit this business problem: A student-support model uses attendance history to flag risk without first examining stigma or false positives.

  1. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  2. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  3. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

C:

Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Review a documented impact assessment with risk owners, affected groups and recorded design responses. Together, they implement and test the intended safeguard. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Incorrect Answers

A:

This pair supports another governance intervention: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

B:

This pair supports another governance intervention: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

D:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Question 396

A clinic designs a resource-priority assistant before assessing how errors could affect vulnerable patients. Which governance action most directly addresses the need in this situation?

  1. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  2. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  3. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  4. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

Explanation

Correct Answer

D:

For the decision, evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. The practical acceptance evidence is: review a documented impact assessment with risk owners, affected groups and recorded design responses

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. The more direct step is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open

B:

The alternative—first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings—would not resolve the stated issue. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

C:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—would not resolve the stated issue. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.

Question 397

The organization chooses to evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open for this situation: A clinic designs a resource-priority assistant before assessing how errors could affect vulnerable patients. Which rationale best supports that choice?

  1. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  2. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.
  3. Design changes and restrictions can prevent harm more effectively than disclosures alone.
  4. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Explanation

Correct Answer

A:

For the technical explanation, early impact analysis can change or reject an AI design before deployment makes risks harder to reduce. That is why the case calls for this action: evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open

Incorrect Answers

B:

That explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

C:

That explains why a team might first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

D:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.

Question 398

The organization has implemented a control for this case: A clinic designs a resource-priority assistant before assessing how errors could affect vulnerable patients. Which result would best validate it?

  1. Validate the use-case specification with business and affected-user representatives before model selection.
  2. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  3. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

B:

For the verification, review a documented impact assessment with risk owners, affected groups and recorded design responses. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce

Incorrect Answers

A:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

C:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

D:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to review a documented impact assessment with risk owners, affected groups and recorded design responses.

Question 399

A team considers this AI use: A clinic designs a resource-priority assistant before assessing how errors could affect vulnerable patients. Rather than applying the relevant control, it will wait to assess foreseeable harms until an AI system has completed its first production cycle. What can go wrong?

  1. Auditors and incident responders cannot recover why critical choices were made.
  2. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  3. Serious issues can become embedded in model features and operating procedures.
  4. Previously approved controls and thresholds may no longer match the deployed model.

Explanation

Correct Answer

C:

For the risk comparison, serious issues can become embedded in model features and operating procedures. The shortcut overlooks the central safeguard: evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

B:

This consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

D:

This consequence follows more directly from another shortcut: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. The shortcut actually described would instead create this risk: Serious issues can become embedded in model features and operating procedures.

Question 400

The governance team needs both a decision and a way to verify it for this case: A clinic designs a resource-priority assistant before assessing how errors could affect vulnerable patients. Which pair is best?

  1. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  2. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.
  3. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  4. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Explanation

Correct Answer

D:

For the action and verification, evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Review a documented impact assessment with risk owners, affected groups and recorded design responses. Together, they implement and test the intended safeguard. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses

B:

This pair supports another governance intervention: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

C:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. The supporting evidence is: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Question 401

A lending analyst receives hundreds of automatic denials per minute with no basis for review. Which step would provide the most targeted control for the problem described?

  1. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  2. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.
  3. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  4. Include privacy, adversarial, access-control and integration tests in training and system evaluation.

Explanation

Correct Answer

A:

Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. A human approval screen is inadequate when reviewers lack time, context or permission to override automation. The practical acceptance evidence is: observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Incorrect Answers

B:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

C:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

D:

This would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

Question 402

A lending analyst receives hundreds of automatic denials per minute with no basis for review. Why is the recommended governance approach appropriate?

  1. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  2. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  3. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.
  4. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Explanation

Correct Answer

B:

A human approval screen is inadequate when reviewers lack time, context or permission to override automation. That is why the case calls for this action: assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

Incorrect Answers

A:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

C:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

D:

That explains why a team might require a model or system card, deployment limits, safety tests and responsible release approval before production, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Question 403

The organization has implemented a control for this case: A lending analyst receives hundreds of automatic denials per minute with no basis for review. Which result would best validate it?

  1. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  2. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

C:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Incorrect Answers

A:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

B:

Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

D:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Question 404

A team considers this AI use: A lending analyst receives hundreds of automatic denials per minute with no basis for review. Rather than applying the relevant control, it will add a nominal approval button while requiring staff to accept results within seconds. What can go wrong?

  1. Serious issues can become embedded in model features and operating procedures.
  2. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  3. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  4. Human involvement becomes a rubber stamp rather than an effective safeguard.

Explanation

Correct Answer

D:

Human involvement becomes a rubber stamp rather than an effective safeguard. The shortcut overlooks the central safeguard: assign a trained reviewer authority, decision information and a realistic opportunity to intervene. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard.

B:

This consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard.

C:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard.

Question 405

A lending analyst receives hundreds of automatic denials per minute with no basis for review. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  2. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  3. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  4. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

A:

Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions. Together, they implement and test the intended safeguard. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Incorrect Answers

B:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

C:

This pair supports another governance intervention: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

D:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Question 406

An AI project team faces this situation: A parole-support tool asks officers to acknowledge risk scores without exposing reasons or override controls. What should the team do first to manage the central issue?

  1. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  2. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  3. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  4. Require notification, evaluation and approval paths for significant updates to externally supplied models.

Explanation

Correct Answer

B:

For the decision, assign a trained reviewer authority, decision information and a realistic opportunity to intervene. A human approval screen is inadequate when reviewers lack time, context or permission to override automation. The practical acceptance evidence is: observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions

Incorrect Answers

A:

This would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

C:

This would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

D:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. The more direct step is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

Question 407

A parole-support tool asks officers to acknowledge risk scores without exposing reasons or override controls. Why is the recommended governance approach appropriate?

  1. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  2. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  3. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  4. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Explanation

Correct Answer

C:

For the technical explanation, a human approval screen is inadequate when reviewers lack time, context or permission to override automation. That is why the case calls for this action: assign a trained reviewer authority, decision information and a realistic opportunity to intervene

Incorrect Answers

A:

For the technical explanation, that explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation

B:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

D:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Question 408

A parole-support tool asks officers to acknowledge risk scores without exposing reasons or override controls. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Validate the use-case specification with business and affected-user representatives before model selection.
  2. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Explanation

Correct Answer

D:

For the verification, observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A human approval screen is inadequate when reviewers lack time, context or permission to override automation

Incorrect Answers

A:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

B:

For the verification, verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions

C:

Review the adversarial test inventory, observed failures, remediations and retest outcomes would test a different outcome. The directly relevant evidence is to observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Question 409

For this case—A parole-support tool asks officers to acknowledge risk scores without exposing reasons or override controls—staff choose the shortcut: Add a nominal approval button while requiring staff to accept results within seconds. Which consequence best explains why this is inadequate?

  1. Human involvement becomes a rubber stamp rather than an effective safeguard.
  2. Known risks may remain uncontrolled despite formal notice to users.
  3. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  4. Production begins without a defensible record of known limits or release acceptance.

Explanation

Correct Answer

A:

For the risk comparison, human involvement becomes a rubber stamp rather than an effective safeguard. The shortcut overlooks the central safeguard: assign a trained reviewer authority, decision information and a realistic opportunity to intervene. A human approval screen is inadequate when reviewers lack time, context or permission to override automation

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard

D:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Human involvement becomes a rubber stamp rather than an effective safeguard.

Question 410

A parole-support tool asks officers to acknowledge risk scores without exposing reasons or override controls. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  2. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  3. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.
  4. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.

Explanation

Correct Answer

B:

For the action and verification, assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions. Together, they implement and test the intended safeguard. A human approval screen is inadequate when reviewers lack time, context or permission to override automation

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

C:

This pair supports another governance intervention: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

D:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. The supporting evidence is: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Question 411

A factory safety vision model rarely misses a worker yet a miss could cause serious injury. Which step would provide the most targeted control for the problem described?

  1. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  2. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  3. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  4. Record data origin, collection context, permitted use, transformations and lineage for training inputs.

Explanation

Correct Answer

C:

Assess AI failure risks using both estimated occurrence and potential magnitude of harm. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The practical acceptance evidence is: verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Incorrect Answers

A:

This would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

B:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

D:

This would help address a separate governance task: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

Question 412

A team reviewing the following situation asks for the most relevant governance principle: A factory safety vision model rarely misses a worker yet a miss could cause serious injury. Which principle best explains the required control?

  1. Design changes and restrictions can prevent harm more effectively than disclosures alone.
  2. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  3. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  4. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Explanation

Correct Answer

D:

An infrequent but catastrophic error may require stronger controls than a common minor inconvenience. That is why the case calls for this action: assess AI failure risks using both estimated occurrence and potential magnitude of harm.

Incorrect Answers

A:

That explains why a team might first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

B:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

C:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Question 413

An AI governance reviewer examines this situation: A factory safety vision model rarely misses a worker yet a miss could cause serious injury. Which acceptance check is most decisive?

  1. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  2. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  3. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  4. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Explanation

Correct Answer

A:

Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Incorrect Answers

B:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

C:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

D:

Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Question 414

For this case—A factory safety vision model rarely misses a worker yet a miss could cause serious injury—staff choose the shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. Which consequence best explains why this is inadequate?

  1. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  2. Low-frequency, high-impact harms can be neglected.
  3. The design can ignore burdens imposed on people who never chose the tool.
  4. Production begins without a defensible record of known limits or release acceptance.

Explanation

Correct Answer

B:

Low-frequency, high-impact harms can be neglected. The shortcut overlooks the central safeguard: assess AI failure risks using both estimated occurrence and potential magnitude of harm. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected.

C:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected.

D:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected.

Question 415

A factory safety vision model rarely misses a worker yet a miss could cause serious injury. Which action and evidence pair most directly addresses the situation?

  1. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.
  2. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  3. Action: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Evidence: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  4. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Explanation

Correct Answer

C:

Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards. Together, they implement and test the intended safeguard. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Incorrect Answers

A:

This pair supports another governance intervention: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

B:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

D:

This pair supports another governance intervention: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Question 416

A consumer chatbot often misspells store hours while a rare unsafe medical suggestion may cause major harm. Which step would provide the most targeted control for the problem described?

  1. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  2. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.
  3. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  4. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

Explanation

Correct Answer

D:

For the decision, assess AI failure risks using both estimated occurrence and potential magnitude of harm. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The practical acceptance evidence is: verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards

Incorrect Answers

A:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

B:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

C:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. The more direct step is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm.

Question 417

The organization chooses to assess AI failure risks using both estimated occurrence and potential magnitude of harm for this situation: A consumer chatbot often misspells store hours while a rare unsafe medical suggestion may cause major harm. Which rationale best supports that choice?

  1. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  2. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.
  3. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  4. A dataset can be large yet systematically miss important subgroups or operating conditions.

Explanation

Correct Answer

A:

For the technical explanation, an infrequent but catastrophic error may require stronger controls than a common minor inconvenience. That is why the case calls for this action: assess AI failure risks using both estimated occurrence and potential magnitude of harm

Incorrect Answers

B:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

C:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

D:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Question 418

The organization has implemented a control for this case: A consumer chatbot often misspells store hours while a rare unsafe medical suggestion may cause major harm. Which result would best validate it?

  1. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Explanation

Correct Answer

B:

For the verification, verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience

Incorrect Answers

A:

For the verification, check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards

C:

Review a documented impact assessment with risk owners, affected groups and recorded design responses would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

D:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Question 419

A team considers this AI use: A consumer chatbot often misspells store hours while a rare unsafe medical suggestion may cause major harm. Rather than applying the relevant control, it will rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. What can go wrong?

  1. Serious issues can become embedded in model features and operating procedures.
  2. The design can ignore burdens imposed on people who never chose the tool.
  3. Low-frequency, high-impact harms can be neglected.
  4. Rights and quality issues cannot be traced to specific sources for correction.

Explanation

Correct Answer

C:

For the risk comparison, low-frequency, high-impact harms can be neglected. The shortcut overlooks the central safeguard: assess AI failure risks using both estimated occurrence and potential magnitude of harm. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected

D:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Low-frequency, high-impact harms can be neglected.

Question 420

A consumer chatbot often misspells store hours while a rare unsafe medical suggestion may cause major harm. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  3. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  4. Action: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Evidence: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Explanation

Correct Answer

D:

For the action and verification, assess AI failure risks using both estimated occurrence and potential magnitude of harm. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards. Together, they implement and test the intended safeguard. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience

Incorrect Answers

A:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

B:

For the action and verification, this pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards

C:

This pair supports another governance intervention: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For the actual case, the required action is: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. The supporting evidence is: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Question 421

A childcare platform proposes automatic welfare action and relies on a disclaimer instead of human authorization. Which governance action most directly addresses the need in this situation?

  1. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  2. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  3. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  4. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Explanation

Correct Answer

A:

First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Design changes and restrictions can prevent harm more effectively than disclosures alone. The practical acceptance evidence is: trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Incorrect Answers

B:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

C:

This would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

D:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

Question 422

A childcare platform proposes automatic welfare action and relies on a disclaimer instead of human authorization. Which explanation best connects the risk to the proposed governance action?

  1. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  2. Design changes and restrictions can prevent harm more effectively than disclosures alone.
  3. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  4. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Explanation

Correct Answer

B:

Design changes and restrictions can prevent harm more effectively than disclosures alone. That is why the case calls for this action: first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

Incorrect Answers

A:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone.

C:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone.

D:

That explains why a team might include privacy, adversarial, access-control and integration tests in training and system evaluation, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone.

Question 423

The organization has implemented a control for this case: A childcare platform proposes automatic welfare action and relies on a disclaimer instead of human authorization. Which result would best validate it?

  1. Validate the use-case specification with business and affected-user representatives before model selection.
  2. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  3. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

C:

Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Design changes and restrictions can prevent harm more effectively than disclosures alone.

Incorrect Answers

A:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

B:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

D:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Question 424

For this case—A childcare platform proposes automatic welfare action and relies on a disclaimer instead of human authorization—staff choose the shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. Which consequence best explains why this is inadequate?

  1. Auditors and incident responders cannot recover why critical choices were made.
  2. Low-frequency, high-impact harms can be neglected.
  3. Rights and quality issues cannot be traced to specific sources for correction.
  4. Known risks may remain uncontrolled despite formal notice to users.

Explanation

Correct Answer

D:

Known risks may remain uncontrolled despite formal notice to users. The shortcut overlooks the central safeguard: first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Design changes and restrictions can prevent harm more effectively than disclosures alone.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users.

B:

This consequence follows more directly from another shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users.

C:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users.

Question 425

The governance team needs both a decision and a way to verify it for this case: A childcare platform proposes automatic welfare action and relies on a disclaimer instead of human authorization. Which pair is best?

  1. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  2. Action: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Evidence: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.

Explanation

Correct Answer

A:

First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance. Together, they implement and test the intended safeguard. Design changes and restrictions can prevent harm more effectively than disclosures alone.

Incorrect Answers

B:

This pair supports another governance intervention: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

C:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

D:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Question 426

Consider the following proposed AI use: An industrial assistant is allowed to change machine settings after merely displaying a safety warning. Which response best addresses the governance gap?

  1. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  2. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  3. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.
  4. Require notification, evaluation and approval paths for significant updates to externally supplied models.

Explanation

Correct Answer

B:

For the decision, first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Design changes and restrictions can prevent harm more effectively than disclosures alone. The practical acceptance evidence is: trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings

C:

This would help address a separate governance task: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

D:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, design changes and restrictions can prevent harm more effectively than disclosures alone. The more direct step is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

Question 427

A team reviewing the following situation asks for the most relevant governance principle: An industrial assistant is allowed to change machine settings after merely displaying a safety warning. Which principle best explains the required control?

  1. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.
  2. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  3. Design changes and restrictions can prevent harm more effectively than disclosures alone.
  4. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Explanation

Correct Answer

C:

For the technical explanation, design changes and restrictions can prevent harm more effectively than disclosures alone. That is why the case calls for this action: first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings

Incorrect Answers

A:

That explains why a team might write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone.

B:

For the technical explanation, that explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone

D:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. Design changes and restrictions can prevent harm more effectively than disclosures alone.

Question 428

An industrial assistant is allowed to change machine settings after merely displaying a safety warning. Which recorded test or observation most directly demonstrates an effective response?

  1. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  3. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  4. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Explanation

Correct Answer

D:

For the verification, trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Design changes and restrictions can prevent harm more effectively than disclosures alone

Incorrect Answers

A:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

B:

For the verification, observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance

C:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Question 429

The organization has this problem: An industrial assistant is allowed to change machine settings after merely displaying a safety warning. It takes a shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. What is the principal risk?

  1. Known risks may remain uncontrolled despite formal notice to users.
  2. Low-frequency, high-impact harms can be neglected.
  3. The solution may optimize an irrelevant metric while exposing people to unassessed harms.
  4. The model can learn labeling errors and reproduce inconsistent operational decisions.

Explanation

Correct Answer

A:

For the risk comparison, known risks may remain uncontrolled despite formal notice to users. The shortcut overlooks the central safeguard: first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Design changes and restrictions can prevent harm more effectively than disclosures alone

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users

C:

This consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users.

D:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Known risks may remain uncontrolled despite formal notice to users.

Question 430

An industrial assistant is allowed to change machine settings after merely displaying a safety warning. Which action and evidence pair most directly addresses the situation?

  1. Action: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. Evidence: Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  2. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  3. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  4. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

B:

For the action and verification, first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance. Together, they implement and test the intended safeguard. Design changes and restrictions can prevent harm more effectively than disclosures alone

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance

C:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

D:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. The supporting evidence is: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Question 431

An AI project team faces this situation: A university changes its admissions model architecture three times but retains no design rationale. What should the team do first to manage the central issue?

  1. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  2. Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open.
  3. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.
  4. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

Explanation

Correct Answer

C:

Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Traceability supports review of how a model was chosen and why residual risks were considered acceptable. The practical acceptance evidence is: reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Incorrect Answers

A:

This would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

B:

This would help address a separate governance task: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

D:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

Question 432

The organization chooses to maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development for this situation: A university changes its admissions model architecture three times but retains no design rationale. Which rationale best supports that choice?

  1. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  2. Design changes and restrictions can prevent harm more effectively than disclosures alone.
  3. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  4. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Explanation

Correct Answer

D:

Traceability supports review of how a model was chosen and why residual risks were considered acceptable. That is why the case calls for this action: maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

Incorrect Answers

A:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

B:

That explains why a team might first consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

C:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Question 433

An AI governance reviewer examines this situation: A university changes its admissions model architecture three times but retains no design rationale. Which acceptance check is most decisive?

  1. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  2. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.
  3. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

A:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Incorrect Answers

B:

The alternative—verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

C:

The alternative—review a documented impact assessment with risk owners, affected groups and recorded design responses—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

D:

The alternative—inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Question 434

A team considers this AI use: A university changes its admissions model architecture three times but retains no design rationale. Rather than applying the relevant control, it will keep only the final model file and discard the records of rejected designs and control decisions. What can go wrong?

  1. The design can ignore burdens imposed on people who never chose the tool.
  2. Auditors and incident responders cannot recover why critical choices were made.
  3. Low-frequency, high-impact harms can be neglected.
  4. Quality can degrade gradually without generating a conspicuous technical outage.

Explanation

Correct Answer

B:

Auditors and incident responders cannot recover why critical choices were made. The shortcut overlooks the central safeguard: maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made.

C:

This consequence follows more directly from another shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made.

D:

This consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made.

Question 435

Choose the intervention and validation that fit this business problem: A university changes its admissions model architecture three times but retains no design rationale.

  1. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  2. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  3. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  4. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Explanation

Correct Answer

C:

Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals. Together, they implement and test the intended safeguard. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Incorrect Answers

A:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

B:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

D:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Question 436

Consider the following proposed AI use: A vendor rotates its data-science team and loses evidence behind feature exclusions. Which response best addresses the governance gap?

  1. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  2. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  3. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  4. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

Explanation

Correct Answer

D:

For the decision, maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Traceability supports review of how a model was chosen and why residual risks were considered acceptable. The practical acceptance evidence is: reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals

Incorrect Answers

A:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

B:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

C:

For the decision, this would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, traceability supports review of how a model was chosen and why residual risks were considered acceptable. The more direct step is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development

Question 437

A vendor rotates its data-science team and loses evidence behind feature exclusions. Why is the recommended governance approach appropriate?

  1. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.
  2. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.
  3. Early impact analysis can change or reject an AI design before deployment makes risks harder to reduce.
  4. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Explanation

Correct Answer

A:

For the technical explanation, traceability supports review of how a model was chosen and why residual risks were considered acceptable. That is why the case calls for this action: maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development

Incorrect Answers

B:

For the technical explanation, that explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable

C:

That explains why a team might evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

D:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Question 438

A vendor rotates its data-science team and loses evidence behind feature exclusions. Which recorded test or observation most directly demonstrates an effective response?

  1. Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  3. Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

B:

For the verification, reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Traceability supports review of how a model was chosen and why residual risks were considered acceptable

Incorrect Answers

A:

The alternative—check stakeholder maps and feedback records for representation of affected groups and mitigations—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

C:

The alternative—trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

D:

The alternative—check the assurance calendar, completed review evidence and tracking of overdue corrective actions—would not resolve the stated issue. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Question 439

A vendor rotates its data-science team and loses evidence behind feature exclusions. The team instead proposes: Keep only the final model file and discard the records of rejected designs and control decisions. Which failure should the reviewer expect most directly?

  1. Low-frequency, high-impact harms can be neglected.
  2. The design can ignore burdens imposed on people who never chose the tool.
  3. Auditors and incident responders cannot recover why critical choices were made.
  4. The model can learn labeling errors and reproduce inconsistent operational decisions.

Explanation

Correct Answer

C:

For the risk comparison, auditors and incident responders cannot recover why critical choices were made. The shortcut overlooks the central safeguard: maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Traceability supports review of how a model was chosen and why residual risks were considered acceptable

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Rank risks only by the number of past complaints while ignoring foreseeable severe outcomes. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made

B:

For the risk comparison, this consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made

D:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Auditors and incident responders cannot recover why critical choices were made.

Question 440

The governance team needs both a decision and a way to verify it for this case: A vendor rotates its data-science team and loses evidence behind feature exclusions. Which pair is best?

  1. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  2. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  3. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  4. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Explanation

Correct Answer

D:

For the action and verification, maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals. Together, they implement and test the intended safeguard. Traceability supports review of how a model was chosen and why residual risks were considered acceptable

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals

B:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

C:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. The supporting evidence is: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Question 441

An AI project team faces this situation: A recruitment model incorporates job-board data with unknown original collection terms. What should the team do first to manage the central issue?

  1. Record data origin, collection context, permitted use, transformations and lineage for training inputs.
  2. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  3. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  4. Version models, prompts and data transformations and define tested rollback criteria.

Explanation

Correct Answer

A:

Record data origin, collection context, permitted use, transformations and lineage for training inputs. Reliable provenance helps assess rights, representativeness and the source of discovered model failures. The practical acceptance evidence is: trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Incorrect Answers

B:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

C:

This would help address a separate governance task: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

D:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

Question 442

A team reviewing the following situation asks for the most relevant governance principle: A recruitment model incorporates job-board data with unknown original collection terms. Which principle best explains the required control?

  1. A dataset can be large yet systematically miss important subgroups or operating conditions.
  2. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  3. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  4. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Explanation

Correct Answer

B:

Reliable provenance helps assess rights, representativeness and the source of discovered model failures. That is why the case calls for this action: record data origin, collection context, permitted use, transformations and lineage for training inputs.

Incorrect Answers

A:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

C:

That explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

D:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Question 443

A recruitment model incorporates job-board data with unknown original collection terms. Which recorded test or observation most directly demonstrates an effective response?

  1. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  2. Compare group and context distributions against intended deployment, then test known underrepresented cases.
  3. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  4. Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Explanation

Correct Answer

C:

Trace selected training records back to licensed or approved origins through recorded preprocessing steps. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Incorrect Answers

A:

Review stratified metrics, stress cases and documented thresholds with accountable reviewers would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

B:

Compare group and context distributions against intended deployment, then test known underrepresented cases would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

D:

Compare deployed behavior and current instructions with approved disclosures and update records as needed would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Question 444

A recruitment model incorporates job-board data with unknown original collection terms. The team instead proposes: Combine anonymous internet files and internal records into one dataset with no source log. Which failure should the reviewer expect most directly?

  1. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  2. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.
  3. The design can ignore burdens imposed on people who never chose the tool.
  4. Rights and quality issues cannot be traced to specific sources for correction.

Explanation

Correct Answer

D:

Rights and quality issues cannot be traced to specific sources for correction. The shortcut overlooks the central safeguard: record data origin, collection context, permitted use, transformations and lineage for training inputs. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

B:

This consequence follows more directly from another shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

C:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

Question 445

The governance team needs both a decision and a way to verify it for this case: A recruitment model incorporates job-board data with unknown original collection terms. Which pair is best?

  1. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  2. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  4. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

A:

Record data origin, collection context, permitted use, transformations and lineage for training inputs. Trace selected training records back to licensed or approved origins through recorded preprocessing steps. Together, they implement and test the intended safeguard. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Incorrect Answers

B:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

C:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

D:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Question 446

Consider the following proposed AI use: An agriculture model merges public imagery and partner farm records without recording their lineage. Which response best addresses the governance gap?

  1. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  2. Record data origin, collection context, permitted use, transformations and lineage for training inputs.
  3. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  4. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

Explanation

Correct Answer

B:

For the decision, record data origin, collection context, permitted use, transformations and lineage for training inputs. Reliable provenance helps assess rights, representativeness and the source of discovered model failures. The practical acceptance evidence is: trace selected training records back to licensed or approved origins through recorded preprocessing steps

Incorrect Answers

A:

This would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

C:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

D:

This would help address a separate governance task: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For this case, however, reliable provenance helps assess rights, representativeness and the source of discovered model failures. The more direct step is: Record data origin, collection context, permitted use, transformations and lineage for training inputs.

Question 447

A team reviewing the following situation asks for the most relevant governance principle: An agriculture model merges public imagery and partner farm records without recording their lineage. Which principle best explains the required control?

  1. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  2. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  3. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  4. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Explanation

Correct Answer

C:

For the technical explanation, reliable provenance helps assess rights, representativeness and the source of discovered model failures. That is why the case calls for this action: record data origin, collection context, permitted use, transformations and lineage for training inputs

Incorrect Answers

A:

For the technical explanation, that explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures

B:

That explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

D:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Question 448

The organization has implemented a control for this case: An agriculture model merges public imagery and partner farm records without recording their lineage. Which result would best validate it?

  1. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  2. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.
  4. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Explanation

Correct Answer

D:

For the verification, trace selected training records back to licensed or approved origins through recorded preprocessing steps. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Reliable provenance helps assess rights, representativeness and the source of discovered model failures

Incorrect Answers

A:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

B:

Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

C:

Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions would test a different outcome. The directly relevant evidence is to trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Question 449

For this case—An agriculture model merges public imagery and partner farm records without recording their lineage—staff choose the shortcut: Combine anonymous internet files and internal records into one dataset with no source log. Which consequence best explains why this is inadequate?

  1. Rights and quality issues cannot be traced to specific sources for correction.
  2. Reported generalization may be overstated because the test set contains known information.
  3. The model can perform badly on neglected populations despite strong headline metrics.
  4. Operators cannot reliably reverse a harmful release or identify what changed.

Explanation

Correct Answer

A:

For the risk comparison, rights and quality issues cannot be traced to specific sources for correction. The shortcut overlooks the central safeguard: record data origin, collection context, permitted use, transformations and lineage for training inputs. Reliable provenance helps assess rights, representativeness and the source of discovered model failures

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Randomly divide rows even when identical customers and future outcomes occur in both sets. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

C:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

D:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Rights and quality issues cannot be traced to specific sources for correction.

Question 450

Choose the intervention and validation that fit this business problem: An agriculture model merges public imagery and partner farm records without recording their lineage.

  1. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  2. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  3. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  4. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Explanation

Correct Answer

B:

For the action and verification, record data origin, collection context, permitted use, transformations and lineage for training inputs. Trace selected training records back to licensed or approved origins through recorded preprocessing steps. Together, they implement and test the intended safeguard. Reliable provenance helps assess rights, representativeness and the source of discovered model failures

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps

C:

This pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

D:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Record data origin, collection context, permitted use, transformations and lineage for training inputs. The supporting evidence is: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.

Question 451

A voice assistant is trained on urban accents but deployed to rural emergency callers. Which governance action most directly addresses the need in this situation?

  1. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  2. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  3. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  4. Version models, prompts and data transformations and define tested rollback criteria.

Explanation

Correct Answer

C:

Evaluate whether training and validation samples cover the populations and conditions the model will encounter. A dataset can be large yet systematically miss important subgroups or operating conditions. The practical acceptance evidence is: compare group and context distributions against intended deployment, then test known underrepresented cases.

Incorrect Answers

A:

This would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

B:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

D:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

Question 452

A team reviewing the following situation asks for the most relevant governance principle: A voice assistant is trained on urban accents but deployed to rural emergency callers. Which principle best explains the required control?

  1. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  2. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  3. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.
  4. A dataset can be large yet systematically miss important subgroups or operating conditions.

Explanation

Correct Answer

D:

A dataset can be large yet systematically miss important subgroups or operating conditions. That is why the case calls for this action: evaluate whether training and validation samples cover the populations and conditions the model will encounter.

Incorrect Answers

A:

That explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions.

B:

That explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions.

C:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions.

Question 453

The organization has implemented a control for this case: A voice assistant is trained on urban accents but deployed to rural emergency callers. Which result would best validate it?

  1. Compare group and context distributions against intended deployment, then test known underrepresented cases.
  2. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  3. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  4. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Explanation

Correct Answer

A:

Compare group and context distributions against intended deployment, then test known underrepresented cases. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A dataset can be large yet systematically miss important subgroups or operating conditions.

Incorrect Answers

B:

Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases.

C:

Review stratified metrics, stress cases and documented thresholds with accountable reviewers would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases.

D:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases.

Question 454

The organization has this problem: A voice assistant is trained on urban accents but deployed to rural emergency callers. It takes a shortcut: Assume millions of data rows guarantee fairness and coverage for every user. What is the principal risk?

  1. The model can learn labeling errors and reproduce inconsistent operational decisions.
  2. The model can perform badly on neglected populations despite strong headline metrics.
  3. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.
  4. Auditors and incident responders cannot recover why critical choices were made.

Explanation

Correct Answer

B:

The model can perform badly on neglected populations despite strong headline metrics. The shortcut overlooks the central safeguard: evaluate whether training and validation samples cover the populations and conditions the model will encounter. A dataset can be large yet systematically miss important subgroups or operating conditions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

C:

This consequence follows more directly from another shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

D:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

Question 455

A voice assistant is trained on urban accents but deployed to rural emergency callers. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  2. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  3. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  4. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

Explanation

Correct Answer

C:

Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Compare group and context distributions against intended deployment, then test known underrepresented cases. Together, they implement and test the intended safeguard. A dataset can be large yet systematically miss important subgroups or operating conditions.

Incorrect Answers

A:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

B:

This pair supports another governance intervention: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

D:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

Question 456

Consider the following proposed AI use: A road hazard model is trained mostly on daylight highways but used on rural night routes. Which response best addresses the governance gap?

  1. Record data origin, collection context, permitted use, transformations and lineage for training inputs.
  2. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  3. Assess AI failure risks using both estimated occurrence and potential magnitude of harm.
  4. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

Explanation

Correct Answer

D:

For the decision, evaluate whether training and validation samples cover the populations and conditions the model will encounter. A dataset can be large yet systematically miss important subgroups or operating conditions. The practical acceptance evidence is: compare group and context distributions against intended deployment, then test known underrepresented cases

Incorrect Answers

A:

This would help address a separate governance task: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

B:

This would help address a separate governance task: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

C:

This would help address a separate governance task: Assess AI failure risks using both estimated occurrence and potential magnitude of harm. For this case, however, a dataset can be large yet systematically miss important subgroups or operating conditions. The more direct step is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter.

Question 457

The organization chooses to evaluate whether training and validation samples cover the populations and conditions the model will encounter for this situation: A road hazard model is trained mostly on daylight highways but used on rural night routes. Which rationale best supports that choice?

  1. A dataset can be large yet systematically miss important subgroups or operating conditions.
  2. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  3. Average benchmark performance alone is not evidence of reliable service for every affected group.
  4. Controls may erode as use cases, personnel, vendors and model dependencies change.

Explanation

Correct Answer

A:

For the technical explanation, a dataset can be large yet systematically miss important subgroups or operating conditions. That is why the case calls for this action: evaluate whether training and validation samples cover the populations and conditions the model will encounter

Incorrect Answers

B:

For the technical explanation, that explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions

C:

That explains why a team might test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions.

D:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. A dataset can be large yet systematically miss important subgroups or operating conditions.

Question 458

A road hazard model is trained mostly on daylight highways but used on rural night routes. Which recorded test or observation most directly demonstrates an effective response?

  1. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  2. Compare group and context distributions against intended deployment, then test known underrepresented cases.
  3. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  4. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

B:

For the verification, compare group and context distributions against intended deployment, then test known underrepresented cases. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A dataset can be large yet systematically miss important subgroups or operating conditions

Incorrect Answers

A:

Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases.

C:

For the verification, inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases

D:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to compare group and context distributions against intended deployment, then test known underrepresented cases.

Question 459

A team considers this AI use: A road hazard model is trained mostly on daylight highways but used on rural night routes. Rather than applying the relevant control, it will assume millions of data rows guarantee fairness and coverage for every user. What can go wrong?

  1. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  2. Reported generalization may be overstated because the test set contains known information.
  3. The model can perform badly on neglected populations despite strong headline metrics.
  4. Previously approved controls and thresholds may no longer match the deployed model.

Explanation

Correct Answer

C:

For the risk comparison, the model can perform badly on neglected populations despite strong headline metrics. The shortcut overlooks the central safeguard: evaluate whether training and validation samples cover the populations and conditions the model will encounter. A dataset can be large yet systematically miss important subgroups or operating conditions

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

B:

This consequence follows more directly from another shortcut: Randomly divide rows even when identical customers and future outcomes occur in both sets. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

D:

This consequence follows more directly from another shortcut: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. The shortcut actually described would instead create this risk: The model can perform badly on neglected populations despite strong headline metrics.

Question 460

A road hazard model is trained mostly on daylight highways but used on rural night routes. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  2. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.
  4. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.

Explanation

Correct Answer

D:

For the action and verification, evaluate whether training and validation samples cover the populations and conditions the model will encounter. Compare group and context distributions against intended deployment, then test known underrepresented cases. Together, they implement and test the intended safeguard. A dataset can be large yet systematically miss important subgroups or operating conditions

Incorrect Answers

A:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

B:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

C:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. The supporting evidence is: Compare group and context distributions against intended deployment, then test known underrepresented cases.

Question 461

A customer-service classifier inherits conflicting issue categories from three support centers. Which step would provide the most targeted control for the problem described?

  1. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  2. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  3. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  4. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

Explanation

Correct Answer

A:

Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The practical acceptance evidence is: sample labeled records across sources and measure disagreement, missingness and remediation results.

Incorrect Answers

B:

This would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

C:

This would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

D:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Question 462

A team reviewing the following situation asks for the most relevant governance principle: A customer-service classifier inherits conflicting issue categories from three support centers. Which principle best explains the required control?

  1. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  2. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  3. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.
  4. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Explanation

Correct Answer

B:

Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. That is why the case calls for this action: audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Incorrect Answers

A:

That explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

C:

That explains why a team might include privacy, adversarial, access-control and integration tests in training and system evaluation, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

D:

That explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Question 463

A customer-service classifier inherits conflicting issue categories from three support centers. Which recorded test or observation most directly demonstrates an effective response?

  1. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  2. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  3. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  4. Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Explanation

Correct Answer

C:

Sample labeled records across sources and measure disagreement, missingness and remediation results. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Incorrect Answers

A:

The alternative—inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

B:

The alternative—trace selected training records back to licensed or approved origins through recorded preprocessing steps—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

D:

The alternative—compare deployed behavior and current instructions with approved disclosures and update records as needed—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Question 464

The organization has this problem: A customer-service classifier inherits conflicting issue categories from three support centers. It takes a shortcut: Treat every inherited annotation as correct because it came from a large historical database. What is the principal risk?

  1. The model can perform badly on neglected populations despite strong headline metrics.
  2. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.
  3. Serious issues can become embedded in model features and operating procedures.
  4. The model can learn labeling errors and reproduce inconsistent operational decisions.

Explanation

Correct Answer

D:

The model can learn labeling errors and reproduce inconsistent operational decisions. The shortcut overlooks the central safeguard: audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions.

B:

This consequence follows more directly from another shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions.

C:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions.

Question 465

A customer-service classifier inherits conflicting issue categories from three support centers. Which action and evidence pair most directly addresses the situation?

  1. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.
  2. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  3. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  4. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

Explanation

Correct Answer

A:

Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Sample labeled records across sources and measure disagreement, missingness and remediation results. Together, they implement and test the intended safeguard. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Incorrect Answers

B:

This pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results.

C:

This pair supports another governance intervention: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results.

D:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results.

Question 466

Consider the following proposed AI use: A hospital prediction dataset has diagnosis codes that changed meaning between years. Which response best addresses the governance gap?

  1. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  2. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  3. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  4. Version models, prompts and data transformations and define tested rollback criteria.

Explanation

Correct Answer

B:

For the decision, audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The practical acceptance evidence is: sample labeled records across sources and measure disagreement, missingness and remediation results

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes

C:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

D:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. The more direct step is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Question 467

A team reviewing the following situation asks for the most relevant governance principle: A hospital prediction dataset has diagnosis codes that changed meaning between years. Which principle best explains the required control?

  1. A dataset can be large yet systematically miss important subgroups or operating conditions.
  2. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  3. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  4. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Explanation

Correct Answer

C:

For the technical explanation, inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality. That is why the case calls for this action: audit missing values, inconsistent labels and measurement errors before trusting training outcomes

Incorrect Answers

A:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

B:

For the technical explanation, that explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality

D:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Question 468

The organization has implemented a control for this case: A hospital prediction dataset has diagnosis codes that changed meaning between years. Which result would best validate it?

  1. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  2. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  4. Sample labeled records across sources and measure disagreement, missingness and remediation results.

Explanation

Correct Answer

D:

For the verification, sample labeled records across sources and measure disagreement, missingness and remediation results. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality

Incorrect Answers

A:

For the verification, the alternative—trace selected training records back to licensed or approved origins through recorded preprocessing steps—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality

B:

The alternative—execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

C:

The alternative—compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed—would not resolve the stated issue. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Question 469

A hospital prediction dataset has diagnosis codes that changed meaning between years. The team instead proposes: Treat every inherited annotation as correct because it came from a large historical database. Which failure should the reviewer expect most directly?

  1. The model can learn labeling errors and reproduce inconsistent operational decisions.
  2. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  3. Rights and quality issues cannot be traced to specific sources for correction.
  4. Serious issues can become embedded in model features and operating procedures.

Explanation

Correct Answer

A:

For the risk comparison, the model can learn labeling errors and reproduce inconsistent operational decisions. The shortcut overlooks the central safeguard: audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions.

C:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: The model can learn labeling errors and reproduce inconsistent operational decisions

Question 470

The governance team needs both a decision and a way to verify it for this case: A hospital prediction dataset has diagnosis codes that changed meaning between years. Which pair is best?

  1. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  2. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.
  3. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  4. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Explanation

Correct Answer

B:

For the action and verification, audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Sample labeled records across sources and measure disagreement, missingness and remediation results. Together, they implement and test the intended safeguard. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results

C:

For the action and verification, this pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results

D:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. The supporting evidence is: Sample labeled records across sources and measure disagreement, missingness and remediation results.

Question 471

A bank forecasts defaults with customer histories that cross both training and test partitions. Which step would provide the most targeted control for the problem described?

  1. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  2. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  3. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  4. Require notification, evaluation and approval paths for significant updates to externally supplied models.

Explanation

Correct Answer

C:

Prevent leakage of future information or duplicate records across training and independent evaluation sets. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The practical acceptance evidence is: inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Incorrect Answers

A:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets.

B:

This would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets.

D:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets.

Question 472

A team reviewing the following situation asks for the most relevant governance principle: A bank forecasts defaults with customer histories that cross both training and test partitions. Which principle best explains the required control?

  1. A dataset can be large yet systematically miss important subgroups or operating conditions.
  2. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.
  3. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  4. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Explanation

Correct Answer

D:

Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. That is why the case calls for this action: prevent leakage of future information or duplicate records across training and independent evaluation sets.

Incorrect Answers

A:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

B:

That explains why a team might include privacy, adversarial, access-control and integration tests in training and system evaluation, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

C:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Question 473

A bank forecasts defaults with customer histories that cross both training and test partitions. Which recorded test or observation most directly demonstrates an effective response?

  1. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  2. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  3. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  4. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Explanation

Correct Answer

A:

Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Incorrect Answers

B:

Review stratified metrics, stress cases and documented thresholds with accountable reviewers would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

C:

Trace selected training records back to licensed or approved origins through recorded preprocessing steps would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

D:

Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Question 474

A team considers this AI use: A bank forecasts defaults with customer histories that cross both training and test partitions. Rather than applying the relevant control, it will randomly divide rows even when identical customers and future outcomes occur in both sets. What can go wrong?

  1. The model can learn labeling errors and reproduce inconsistent operational decisions.
  2. Reported generalization may be overstated because the test set contains known information.
  3. Rights and quality issues cannot be traced to specific sources for correction.
  4. Serious issues can become embedded in model features and operating procedures.

Explanation

Correct Answer

B:

Reported generalization may be overstated because the test set contains known information. The shortcut overlooks the central safeguard: prevent leakage of future information or duplicate records across training and independent evaluation sets. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

C:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

D:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

Question 475

A bank forecasts defaults with customer histories that cross both training and test partitions. Which action and evidence pair most directly addresses the situation?

  1. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  2. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  3. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  4. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Explanation

Correct Answer

C:

Prevent leakage of future information or duplicate records across training and independent evaluation sets. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population. Together, they implement and test the intended safeguard. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Incorrect Answers

A:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

B:

This pair supports another governance intervention: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

D:

This pair supports another governance intervention: Set post-release alerts for changes in input populations, outcomes and model quality. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Question 476

Consider the following proposed AI use: A warehouse forecasts demand using future replenishment data accidentally available during model fitting. Which response best addresses the governance gap?

  1. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  2. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  3. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  4. Prevent leakage of future information or duplicate records across training and independent evaluation sets.

Explanation

Correct Answer

D:

For the decision, prevent leakage of future information or duplicate records across training and independent evaluation sets. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The practical acceptance evidence is: inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets

B:

This would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets.

C:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. The more direct step is: Prevent leakage of future information or duplicate records across training and independent evaluation sets.

Question 477

A team reviewing the following situation asks for the most relevant governance principle: A warehouse forecasts demand using future replenishment data accidentally available during model fitting. Which principle best explains the required control?

  1. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  2. A dataset can be large yet systematically miss important subgroups or operating conditions.
  3. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  4. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Explanation

Correct Answer

A:

For the technical explanation, leakage can make a model appear accurate on paper without predicting genuinely unseen conditions. That is why the case calls for this action: prevent leakage of future information or duplicate records across training and independent evaluation sets

Incorrect Answers

B:

For the technical explanation, that explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions

C:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

D:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.

Question 478

A warehouse forecasts demand using future replenishment data accidentally available during model fitting. Which recorded test or observation most directly demonstrates an effective response?

  1. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  2. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  3. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  4. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Explanation

Correct Answer

B:

For the verification, inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions

Incorrect Answers

A:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

C:

For the verification, review stratified metrics, stress cases and documented thresholds with accountable reviewers would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population

D:

For the verification, verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards would test a different outcome. The directly relevant evidence is to inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population

Question 479

A warehouse forecasts demand using future replenishment data accidentally available during model fitting. The team instead proposes: Randomly divide rows even when identical customers and future outcomes occur in both sets. Which failure should the reviewer expect most directly?

  1. The model can perform badly on neglected populations despite strong headline metrics.
  2. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  3. Reported generalization may be overstated because the test set contains known information.
  4. Recurring systemic causes may survive the immediate patch and harm additional users.

Explanation

Correct Answer

C:

For the risk comparison, reported generalization may be overstated because the test set contains known information. The shortcut overlooks the central safeguard: prevent leakage of future information or duplicate records across training and independent evaluation sets. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

B:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

D:

This consequence follows more directly from another shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. The shortcut actually described would instead create this risk: Reported generalization may be overstated because the test set contains known information.

Question 480

A warehouse forecasts demand using future replenishment data accidentally available during model fitting. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.
  2. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  3. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  4. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

D:

For the action and verification, prevent leakage of future information or duplicate records across training and independent evaluation sets. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population. Together, they implement and test the intended safeguard. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions

Incorrect Answers

A:

This pair supports another governance intervention: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

B:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

C:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Prevent leakage of future information or duplicate records across training and independent evaluation sets. The supporting evidence is: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Question 481

Consider the following proposed AI use: A triage model is accurate overall but struggles with uncommon presentations. Which response best addresses the governance gap?

  1. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  2. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  3. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  4. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

Explanation

Correct Answer

A:

Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Average benchmark performance alone is not evidence of reliable service for every affected group. The practical acceptance evidence is: review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Incorrect Answers

B:

This would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

C:

This would help address a separate governance task: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

D:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

Question 482

A triage model is accurate overall but struggles with uncommon presentations. Why is the recommended governance approach appropriate?

  1. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  2. Average benchmark performance alone is not evidence of reliable service for every affected group.
  3. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  4. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Explanation

Correct Answer

B:

Average benchmark performance alone is not evidence of reliable service for every affected group. That is why the case calls for this action: test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

Incorrect Answers

A:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group.

C:

That explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group.

D:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group.

Question 483

The organization has implemented a control for this case: A triage model is accurate overall but struggles with uncommon presentations. Which result would best validate it?

  1. Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  2. Compare group and context distributions against intended deployment, then test known underrepresented cases.
  3. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  4. Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Explanation

Correct Answer

C:

Review stratified metrics, stress cases and documented thresholds with accountable reviewers. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Average benchmark performance alone is not evidence of reliable service for every affected group.

Incorrect Answers

A:

Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

B:

Compare group and context distributions against intended deployment, then test known underrepresented cases would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

D:

Review a documented impact assessment with risk owners, affected groups and recorded design responses would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Question 484

The organization has this problem: A triage model is accurate overall but struggles with uncommon presentations. It takes a shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. What is the principal risk?

  1. The model can learn labeling errors and reproduce inconsistent operational decisions.
  2. Rights and quality issues cannot be traced to specific sources for correction.
  3. The design can ignore burdens imposed on people who never chose the tool.
  4. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

Explanation

Correct Answer

D:

Deployment can expose pockets of high error and brittle behavior not seen in aggregate results. The shortcut overlooks the central safeguard: test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Average benchmark performance alone is not evidence of reliable service for every affected group.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

B:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

C:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

Question 485

Choose the intervention and validation that fit this business problem: A triage model is accurate overall but struggles with uncommon presentations.

  1. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  2. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.
  4. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Explanation

Correct Answer

A:

Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Review stratified metrics, stress cases and documented thresholds with accountable reviewers. Together, they implement and test the intended safeguard. Average benchmark performance alone is not evidence of reliable service for every affected group.

Incorrect Answers

B:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

C:

This pair supports another governance intervention: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

D:

This pair supports another governance intervention: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Question 486

Consider the following proposed AI use: A package-reading vision model fails on damaged labels despite excellent results on clean scans. Which response best addresses the governance gap?

  1. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  2. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  3. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  4. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

Explanation

Correct Answer

B:

For the decision, test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Average benchmark performance alone is not evidence of reliable service for every affected group. The practical acceptance evidence is: review stratified metrics, stress cases and documented thresholds with accountable reviewers

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations

C:

For the decision, this would help address a separate governance task: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations

D:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, average benchmark performance alone is not evidence of reliable service for every affected group. The more direct step is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

Question 487

The organization chooses to test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations for this situation: A package-reading vision model fails on damaged labels despite excellent results on clean scans. Which rationale best supports that choice?

  1. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  2. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.
  3. Average benchmark performance alone is not evidence of reliable service for every affected group.
  4. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Explanation

Correct Answer

C:

For the technical explanation, average benchmark performance alone is not evidence of reliable service for every affected group. That is why the case calls for this action: test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations

Incorrect Answers

A:

That explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group.

B:

For the technical explanation, that explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group

D:

That explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. Average benchmark performance alone is not evidence of reliable service for every affected group.

Question 488

An AI governance reviewer examines this situation: A package-reading vision model fails on damaged labels despite excellent results on clean scans. Which acceptance check is most decisive?

  1. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  2. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  3. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  4. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Explanation

Correct Answer

D:

For the verification, review stratified metrics, stress cases and documented thresholds with accountable reviewers. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Average benchmark performance alone is not evidence of reliable service for every affected group

Incorrect Answers

A:

Trace selected training records back to licensed or approved origins through recorded preprocessing steps would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

B:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

C:

Compare deployed behavior and current instructions with approved disclosures and update records as needed would test a different outcome. The directly relevant evidence is to review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Question 489

The organization has this problem: A package-reading vision model fails on damaged labels despite excellent results on clean scans. It takes a shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. What is the principal risk?

  1. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.
  2. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  3. The model can learn labeling errors and reproduce inconsistent operational decisions.
  4. Operators cannot reliably reverse a harmful release or identify what changed.

Explanation

Correct Answer

A:

For the risk comparison, deployment can expose pockets of high error and brittle behavior not seen in aggregate results. The shortcut overlooks the central safeguard: test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Average benchmark performance alone is not evidence of reliable service for every affected group

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Declare all validation complete after comparing precision and recall alone. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

C:

For the risk comparison, this consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results

D:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.

Question 490

A package-reading vision model fails on damaged labels despite excellent results on clean scans. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.
  2. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  3. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  4. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

B:

For the action and verification, test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Review stratified metrics, stress cases and documented thresholds with accountable reviewers. Together, they implement and test the intended safeguard. Average benchmark performance alone is not evidence of reliable service for every affected group

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers

C:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

D:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. The supporting evidence is: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Question 491

Consider the following proposed AI use: A financial assistant can search internal records and is vulnerable to instructions embedded in retrieved text. Which response best addresses the governance gap?

  1. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.
  2. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  3. Include privacy, adversarial, access-control and integration tests in training and system evaluation.
  4. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

Explanation

Correct Answer

C:

Include privacy, adversarial, access-control and integration tests in training and system evaluation. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The practical acceptance evidence is: execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Incorrect Answers

A:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation.

B:

This would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation.

D:

This would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation.

Question 492

A financial assistant can search internal records and is vulnerable to instructions embedded in retrieved text. Why is the recommended governance approach appropriate?

  1. Leakage can make a model appear accurate on paper without predicting genuinely unseen conditions.
  2. A dataset can be large yet systematically miss important subgroups or operating conditions.
  3. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.
  4. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Explanation

Correct Answer

D:

A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. That is why the case calls for this action: include privacy, adversarial, access-control and integration tests in training and system evaluation.

Incorrect Answers

A:

That explains why a team might prevent leakage of future information or duplicate records across training and independent evaluation sets, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

B:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

C:

That explains why a team might require a model or system card, deployment limits, safety tests and responsible release approval before production, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Question 493

An AI governance reviewer examines this situation: A financial assistant can search internal records and is vulnerable to instructions embedded in retrieved text. Which acceptance check is most decisive?

  1. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  2. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  3. Sample labeled records across sources and measure disagreement, missingness and remediation results.
  4. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Explanation

Correct Answer

A:

Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Incorrect Answers

B:

The alternative—trace selected training records back to licensed or approved origins through recorded preprocessing steps—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

C:

The alternative—sample labeled records across sources and measure disagreement, missingness and remediation results—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

D:

The alternative—verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Question 494

For this case—A financial assistant can search internal records and is vulnerable to instructions embedded in retrieved text—staff choose the shortcut: Declare all validation complete after comparing precision and recall alone. Which consequence best explains why this is inadequate?

  1. Rights and quality issues cannot be traced to specific sources for correction.
  2. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  3. The model can learn labeling errors and reproduce inconsistent operational decisions.
  4. The design can ignore burdens imposed on people who never chose the tool.

Explanation

Correct Answer

B:

Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable. The shortcut overlooks the central safeguard: include privacy, adversarial, access-control and integration tests in training and system evaluation. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

C:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

D:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

Question 495

A financial assistant can search internal records and is vulnerable to instructions embedded in retrieved text. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  2. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.
  3. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  4. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

C:

Include privacy, adversarial, access-control and integration tests in training and system evaluation. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios. Together, they implement and test the intended safeguard. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

B:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

D:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Question 496

A clinical summarizer passes factual tests but may disclose another patient's record under crafted prompts. Which step would provide the most targeted control for the problem described?

  1. Evaluate whether training and validation samples cover the populations and conditions the model will encounter.
  2. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.
  3. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  4. Include privacy, adversarial, access-control and integration tests in training and system evaluation.

Explanation

Correct Answer

D:

For the decision, include privacy, adversarial, access-control and integration tests in training and system evaluation. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The practical acceptance evidence is: execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation

B:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation.

C:

This would help address a separate governance task: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For this case, however, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. The more direct step is: Include privacy, adversarial, access-control and integration tests in training and system evaluation.

Question 497

A clinical summarizer passes factual tests but may disclose another patient's record under crafted prompts. Which explanation best connects the risk to the proposed governance action?

  1. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.
  2. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.
  3. Average benchmark performance alone is not evidence of reliable service for every affected group.
  4. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Explanation

Correct Answer

A:

For the technical explanation, a model that predicts well can still leak secrets, be manipulated or connect unsafely to other services. That is why the case calls for this action: include privacy, adversarial, access-control and integration tests in training and system evaluation

Incorrect Answers

B:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

C:

That explains why a team might test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

D:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Question 498

The organization has implemented a control for this case: A clinical summarizer passes factual tests but may disclose another patient's record under crafted prompts. Which result would best validate it?

  1. Review stratified metrics, stress cases and documented thresholds with accountable reviewers.
  2. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  3. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  4. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Explanation

Correct Answer

B:

For the verification, execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services

Incorrect Answers

A:

The alternative—review stratified metrics, stress cases and documented thresholds with accountable reviewers—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

C:

For the verification, the alternative—trace selected training records back to licensed or approved origins through recorded preprocessing steps—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services

D:

The alternative—confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions—would not resolve the stated issue. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.

Question 499

A clinical summarizer passes factual tests but may disclose another patient's record under crafted prompts. The team instead proposes: Declare all validation complete after comparing precision and recall alone. Which failure should the reviewer expect most directly?

  1. Reported generalization may be overstated because the test set contains known information.
  2. The model can perform badly on neglected populations despite strong headline metrics.
  3. Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.
  4. Quality can degrade gradually without generating a conspicuous technical outage.

Explanation

Correct Answer

C:

For the risk comparison, security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable. The shortcut overlooks the central safeguard: include privacy, adversarial, access-control and integration tests in training and system evaluation. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Randomly divide rows even when identical customers and future outcomes occur in both sets. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

B:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

D:

This consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Security vulnerabilities and privacy leaks remain untested even when predictive performance is acceptable.

Question 500

The governance team needs both a decision and a way to verify it for this case: A clinical summarizer passes factual tests but may disclose another patient's record under crafted prompts. Which pair is best?

  1. Action: Record data origin, collection context, permitted use, transformations and lineage for training inputs. Evidence: Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  2. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.
  3. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  4. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Explanation

Correct Answer

D:

For the action and verification, include privacy, adversarial, access-control and integration tests in training and system evaluation. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios. Together, they implement and test the intended safeguard. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services

Incorrect Answers

A:

This pair supports another governance intervention: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

B:

For the action and verification, this pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios

C:

This pair supports another governance intervention: Require notification, evaluation and approval paths for significant updates to externally supplied models. For the actual case, the required action is: Include privacy, adversarial, access-control and integration tests in training and system evaluation. The supporting evidence is: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Question 501

Consider the following proposed AI use: A local authority wants to launch an AI complaints sorter after a five-minute demonstration. Which response best addresses the governance gap?

  1. Require a model or system card, deployment limits, safety tests and responsible release approval before production.
  2. Set post-release alerts for changes in input populations, outcomes and model quality.
  3. Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.
  4. Record data origin, collection context, permitted use, transformations and lineage for training inputs.

Explanation

Correct Answer

A:

Require a model or system card, deployment limits, safety tests and responsible release approval before production. A reproducible release package connects intended use and tested limitations to an accountable deployment decision. The practical acceptance evidence is: inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Incorrect Answers

B:

This would help address a separate governance task: Set post-release alerts for changes in input populations, outcomes and model quality. For this case, however, a reproducible release package connects intended use and tested limitations to an accountable deployment decision. The more direct step is: Require a model or system card, deployment limits, safety tests and responsible release approval before production.

C:

The alternative—run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate—is not the most direct response to this case. The matching action is to require a model or system card, deployment limits, safety tests and responsible release approval before production.

D:

This would help address a separate governance task: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For this case, however, a reproducible release package connects intended use and tested limitations to an accountable deployment decision. The more direct step is: Require a model or system card, deployment limits, safety tests and responsible release approval before production.

Question 502

A local authority wants to launch an AI complaints sorter after a five-minute demonstration. Which explanation best connects the risk to the proposed governance action?

  1. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  2. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.
  3. Transparency materials must reflect the actual system version and relevant role or legal requirements.
  4. A human approval screen is inadequate when reviewers lack time, context or permission to override automation.

Explanation

Correct Answer

B:

A reproducible release package connects intended use and tested limitations to an accountable deployment decision. That is why the case calls for this action: require a model or system card, deployment limits, safety tests and responsible release approval before production.

Incorrect Answers

A:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

C:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

D:

That explains why a team might assign a trained reviewer authority, decision information and a realistic opportunity to intervene, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Question 503

An AI governance reviewer examines this situation: A local authority wants to launch an AI complaints sorter after a five-minute demonstration. Which acceptance check is most decisive?

  1. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  4. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Explanation

Correct Answer

C:

Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Incorrect Answers

A:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

B:

Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

D:

Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Question 504

The organization has this problem: A local authority wants to launch an AI complaints sorter after a five-minute demonstration. It takes a shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. What is the principal risk?

  1. Previously approved controls and thresholds may no longer match the deployed model.
  2. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  3. Deployment can expose pockets of high error and brittle behavior not seen in aggregate results.
  4. Production begins without a defensible record of known limits or release acceptance.

Explanation

Correct Answer

D:

Production begins without a defensible record of known limits or release acceptance. The shortcut overlooks the central safeguard: require a model or system card, deployment limits, safety tests and responsible release approval before production. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

B:

This consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

C:

This consequence follows more directly from another shortcut: Approve the model from one overall F1 score without testing rare but consequential cases. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

Question 505

A local authority wants to launch an AI complaints sorter after a five-minute demonstration. Which action and evidence pair most directly addresses the situation?

  1. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  3. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  4. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Explanation

Correct Answer

A:

Require a model or system card, deployment limits, safety tests and responsible release approval before production. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner. Together, they implement and test the intended safeguard. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Incorrect Answers

B:

This pair supports another governance intervention: Set post-release alerts for changes in input populations, outcomes and model quality. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

C:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

D:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Question 506

A payment company plans to release a new fraud score before publishing its model card and test summary. Which governance action most directly addresses the need in this situation?

  1. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  2. Require a model or system card, deployment limits, safety tests and responsible release approval before production.
  3. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.
  4. Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations.

Explanation

Correct Answer

B:

For the decision, require a model or system card, deployment limits, safety tests and responsible release approval before production. A reproducible release package connects intended use and tested limitations to an accountable deployment decision. The practical acceptance evidence is: inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner

Incorrect Answers

A:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—is not the most direct response to this case. The matching action is to require a model or system card, deployment limits, safety tests and responsible release approval before production.

C:

This would help address a separate governance task: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For this case, however, a reproducible release package connects intended use and tested limitations to an accountable deployment decision. The more direct step is: Require a model or system card, deployment limits, safety tests and responsible release approval before production.

D:

This would help address a separate governance task: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For this case, however, a reproducible release package connects intended use and tested limitations to an accountable deployment decision. The more direct step is: Require a model or system card, deployment limits, safety tests and responsible release approval before production.

Question 507

A team reviewing the following situation asks for the most relevant governance principle: A payment company plans to release a new fraud score before publishing its model card and test summary. Which principle best explains the required control?

  1. Controls may erode as use cases, personnel, vendors and model dependencies change.
  2. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  3. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.
  4. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Explanation

Correct Answer

C:

For the technical explanation, a reproducible release package connects intended use and tested limitations to an accountable deployment decision. That is why the case calls for this action: require a model or system card, deployment limits, safety tests and responsible release approval before production

Incorrect Answers

A:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

B:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

D:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.

Question 508

A payment company plans to release a new fraud score before publishing its model card and test summary. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Review a documented impact assessment with risk owners, affected groups and recorded design responses.
  4. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Explanation

Correct Answer

D:

For the verification, inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A reproducible release package connects intended use and tested limitations to an accountable deployment decision

Incorrect Answers

A:

For the verification, check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner

B:

Review the adversarial test inventory, observed failures, remediations and retest outcomes would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

C:

Review a documented impact assessment with risk owners, affected groups and recorded design responses would test a different outcome. The directly relevant evidence is to inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Question 509

A payment company plans to release a new fraud score before publishing its model card and test summary. The team instead proposes: Promote a model based only on a positive stakeholder demo with no versioned evidence. Which failure should the reviewer expect most directly?

  1. Production begins without a defensible record of known limits or release acceptance.
  2. Operators cannot reliably reverse a harmful release or identify what changed.
  3. Subsequent drift or control failures can remain unnoticed.
  4. Serious issues can become embedded in model features and operating procedures.

Explanation

Correct Answer

A:

For the risk comparison, production begins without a defensible record of known limits or release acceptance. The shortcut overlooks the central safeguard: require a model or system card, deployment limits, safety tests and responsible release approval before production. A reproducible release package connects intended use and tested limitations to an accountable deployment decision

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

C:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

D:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: Production begins without a defensible record of known limits or release acceptance.

Question 510

Choose the intervention and validation that fit this business problem: A payment company plans to release a new fraud score before publishing its model card and test summary.

  1. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  2. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  3. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Action: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. Evidence: Compare group and context distributions against intended deployment, then test known underrepresented cases.

Explanation

Correct Answer

B:

For the action and verification, require a model or system card, deployment limits, safety tests and responsible release approval before production. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner. Together, they implement and test the intended safeguard. A reproducible release package connects intended use and tested limitations to an accountable deployment decision

Incorrect Answers

A:

This pair supports another governance intervention: Require notification, evaluation and approval paths for significant updates to externally supplied models. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

C:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

D:

This pair supports another governance intervention: Evaluate whether training and validation samples cover the populations and conditions the model will encounter. For the actual case, the required action is: Require a model or system card, deployment limits, safety tests and responsible release approval before production. The supporting evidence is: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.

Question 511

A student-support bot is tested only with polite questions despite access to sensitive records. Which step would provide the most targeted control for the problem described?

  1. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.
  2. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  3. Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.
  4. Record data origin, collection context, permitted use, transformations and lineage for training inputs.

Explanation

Correct Answer

C:

Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The practical acceptance evidence is: review the adversarial test inventory, observed failures, remediations and retest outcomes.

Incorrect Answers

A:

This would help address a separate governance task: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For this case, however, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The more direct step is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

B:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—would not resolve the stated issue. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

D:

This would help address a separate governance task: Record data origin, collection context, permitted use, transformations and lineage for training inputs. For this case, however, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The more direct step is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

Question 512

A student-support bot is tested only with polite questions despite access to sensitive records. Why is the recommended governance approach appropriate?

  1. Controls may erode as use cases, personnel, vendors and model dependencies change.
  2. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.
  3. A model that predicts well can still leak secrets, be manipulated or connect unsafely to other services.
  4. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Explanation

Correct Answer

D:

Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. That is why the case calls for this action: run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

Incorrect Answers

A:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

B:

That explains why a team might require a model or system card, deployment limits, safety tests and responsible release approval before production, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

C:

That explains why a team might include privacy, adversarial, access-control and integration tests in training and system evaluation, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Question 513

A student-support bot is tested only with polite questions despite access to sensitive records. Which recorded test or observation most directly demonstrates an effective response?

  1. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  2. Simulate rollback to a prior approved version and verify restored outputs and access policy.
  3. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  4. Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

A:

Review the adversarial test inventory, observed failures, remediations and retest outcomes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Incorrect Answers

B:

Simulate rollback to a prior approved version and verify restored outputs and access policy would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes.

C:

Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes.

D:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes.

Question 514

A student-support bot is tested only with polite questions despite access to sensitive records. The team instead proposes: Conclude an assistant is safe after ten ordinary successful user requests. Which failure should the reviewer expect most directly?

  1. Production begins without a defensible record of known limits or release acceptance.
  2. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  3. Subsequent drift or control failures can remain unnoticed.
  4. Rights and quality issues cannot be traced to specific sources for correction.

Explanation

Correct Answer

B:

Prompt abuse, unsafe escalation and policy bypass may remain undetected. The shortcut overlooks the central safeguard: run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected.

C:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected.

D:

This consequence follows more directly from another shortcut: Combine anonymous internet files and internal records into one dataset with no source log. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected.

Question 515

A student-support bot is tested only with polite questions despite access to sensitive records. Which action and evidence pair most directly addresses the situation?

  1. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.
  3. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Explanation

Correct Answer

C:

Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Review the adversarial test inventory, observed failures, remediations and retest outcomes. Together, they implement and test the intended safeguard. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Incorrect Answers

A:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

B:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

D:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

Question 516

An AI project team faces this situation: A vendor agent can create purchase orders but has not been tested against malicious tool instructions. What should the team do first to manage the central issue?

  1. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.
  2. Version models, prompts and data transformations and define tested rollback criteria.
  3. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.
  4. Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

Explanation

Correct Answer

D:

For the decision, run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The practical acceptance evidence is: review the adversarial test inventory, observed failures, remediations and retest outcomes

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For this case, however, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The more direct step is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate

B:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The more direct step is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

C:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. The more direct step is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.

Question 517

A team reviewing the following situation asks for the most relevant governance principle: A vendor agent can create purchase orders but has not been tested against malicious tool instructions. Which principle best explains the required control?

  1. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.
  2. Real-world distributions can change after launch, making old validation evidence insufficient.
  3. Controls may erode as use cases, personnel, vendors and model dependencies change.
  4. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Explanation

Correct Answer

A:

For the technical explanation, friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths. That is why the case calls for this action: run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate

Incorrect Answers

B:

That explains why a team might set post-release alerts for changes in input populations, outcomes and model quality, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

C:

For the technical explanation, that explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths

D:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.

Question 518

A vendor agent can create purchase orders but has not been tested against malicious tool instructions. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  2. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  4. Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

B:

For the verification, review the adversarial test inventory, observed failures, remediations and retest outcomes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths

Incorrect Answers

A:

Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes.

C:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes.

D:

For the verification, check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to review the adversarial test inventory, observed failures, remediations and retest outcomes

Question 519

For this case—A vendor agent can create purchase orders but has not been tested against malicious tool instructions—staff choose the shortcut: Conclude an assistant is safe after ten ordinary successful user requests. Which consequence best explains why this is inadequate?

  1. Recurring systemic causes may survive the immediate patch and harm additional users.
  2. Production begins without a defensible record of known limits or release acceptance.
  3. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  4. Serious issues can become embedded in model features and operating procedures.

Explanation

Correct Answer

C:

For the risk comparison, prompt abuse, unsafe escalation and policy bypass may remain undetected. The shortcut overlooks the central safeguard: run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected

D:

This consequence follows more directly from another shortcut: Wait to assess foreseeable harms until an AI system has completed its first production cycle. The shortcut actually described would instead create this risk: Prompt abuse, unsafe escalation and policy bypass may remain undetected.

Question 520

The governance team needs both a decision and a way to verify it for this case: A vendor agent can create purchase orders but has not been tested against malicious tool instructions. Which pair is best?

  1. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  2. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  3. Action: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. Evidence: Check stakeholder maps and feedback records for representation of affected groups and mitigations.
  4. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

Explanation

Correct Answer

D:

For the action and verification, run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Review the adversarial test inventory, observed failures, remediations and retest outcomes. Together, they implement and test the intended safeguard. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths

Incorrect Answers

A:

This pair supports another governance intervention: Require notification, evaluation and approval paths for significant updates to externally supplied models. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

B:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

C:

This pair supports another governance intervention: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For the actual case, the required action is: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. The supporting evidence is: Review the adversarial test inventory, observed failures, remediations and retest outcomes.

Question 521

A seasonal demand model's predictions worsen after a major shift in purchasing behavior. Which governance action most directly addresses the need in this situation?

  1. Set post-release alerts for changes in input populations, outcomes and model quality.
  2. Version models, prompts and data transformations and define tested rollback criteria.
  3. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  4. Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development.

Explanation

Correct Answer

A:

Set post-release alerts for changes in input populations, outcomes and model quality. Real-world distributions can change after launch, making old validation evidence insufficient. The practical acceptance evidence is: compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Incorrect Answers

B:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality.

C:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality.

D:

This would help address a separate governance task: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality.

Question 522

A team reviewing the following situation asks for the most relevant governance principle: A seasonal demand model's predictions worsen after a major shift in purchasing behavior. Which principle best explains the required control?

  1. Transparency materials must reflect the actual system version and relevant role or legal requirements.
  2. Real-world distributions can change after launch, making old validation evidence insufficient.
  3. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  4. Inaccurate or inconsistent labels can teach the model an unstable target regardless of algorithm quality.

Explanation

Correct Answer

B:

Real-world distributions can change after launch, making old validation evidence insufficient. That is why the case calls for this action: set post-release alerts for changes in input populations, outcomes and model quality.

Incorrect Answers

A:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient.

C:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient.

D:

That explains why a team might audit missing values, inconsistent labels and measurement errors before trusting training outcomes, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient.

Question 523

The organization has implemented a control for this case: A seasonal demand model's predictions worsen after a major shift in purchasing behavior. Which result would best validate it?

  1. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  3. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  4. Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

C:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Real-world distributions can change after launch, making old validation evidence insufficient.

Incorrect Answers

A:

Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

B:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

D:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Question 524

For this case—A seasonal demand model's predictions worsen after a major shift in purchasing behavior—staff choose the shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. Which consequence best explains why this is inadequate?

  1. Subsequent drift or control failures can remain unnoticed.
  2. Production begins without a defensible record of known limits or release acceptance.
  3. The design can ignore burdens imposed on people who never chose the tool.
  4. Quality can degrade gradually without generating a conspicuous technical outage.

Explanation

Correct Answer

D:

Quality can degrade gradually without generating a conspicuous technical outage. The shortcut overlooks the central safeguard: set post-release alerts for changes in input populations, outcomes and model quality. Real-world distributions can change after launch, making old validation evidence insufficient.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

B:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

C:

This consequence follows more directly from another shortcut: Consult only the purchaser because it funds the AI deployment. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

Question 525

The governance team needs both a decision and a way to verify it for this case: A seasonal demand model's predictions worsen after a major shift in purchasing behavior. Which pair is best?

  1. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  2. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  3. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Action: Prevent leakage of future information or duplicate records across training and independent evaluation sets. Evidence: Inspect split methodology and duplicated entities, then compare performance on a truly held-out period or population.

Explanation

Correct Answer

A:

Set post-release alerts for changes in input populations, outcomes and model quality. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed. Together, they implement and test the intended safeguard. Real-world distributions can change after launch, making old validation evidence insufficient.

Incorrect Answers

B:

This pair supports another governance intervention: Require notification, evaluation and approval paths for significant updates to externally supplied models. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

C:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

D:

This pair supports another governance intervention: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Question 526

A flood warning model begins receiving readings from a replacement sensor network. Which governance action most directly addresses the need in this situation?

  1. Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate.
  2. Set post-release alerts for changes in input populations, outcomes and model quality.
  3. Version models, prompts and data transformations and define tested rollback criteria.
  4. Assign a trained reviewer authority, decision information and a realistic opportunity to intervene.

Explanation

Correct Answer

B:

For the decision, set post-release alerts for changes in input populations, outcomes and model quality. Real-world distributions can change after launch, making old validation evidence insufficient. The practical acceptance evidence is: compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed

Incorrect Answers

A:

This would help address a separate governance task: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality.

C:

For the decision, this would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality

D:

This would help address a separate governance task: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For this case, however, real-world distributions can change after launch, making old validation evidence insufficient. The more direct step is: Set post-release alerts for changes in input populations, outcomes and model quality.

Question 527

A team reviewing the following situation asks for the most relevant governance principle: A flood warning model begins receiving readings from a replacement sensor network. Which principle best explains the required control?

  1. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.
  2. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  3. Real-world distributions can change after launch, making old validation evidence insufficient.
  4. A dataset can be large yet systematically miss important subgroups or operating conditions.

Explanation

Correct Answer

C:

For the technical explanation, real-world distributions can change after launch, making old validation evidence insufficient. That is why the case calls for this action: set post-release alerts for changes in input populations, outcomes and model quality

Incorrect Answers

A:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient.

B:

For the technical explanation, that explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient

D:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Real-world distributions can change after launch, making old validation evidence insufficient.

Question 528

An AI governance reviewer examines this situation: A flood warning model begins receiving readings from a replacement sensor network. Which acceptance check is most decisive?

  1. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  2. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  3. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  4. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Explanation

Correct Answer

D:

For the verification, compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Real-world distributions can change after launch, making old validation evidence insufficient

Incorrect Answers

A:

Compare deployed behavior and current instructions with approved disclosures and update records as needed would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

B:

Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

C:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Question 529

For this case—A flood warning model begins receiving readings from a replacement sensor network—staff choose the shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. Which consequence best explains why this is inadequate?

  1. Quality can degrade gradually without generating a conspicuous technical outage.
  2. Recurring systemic causes may survive the immediate patch and harm additional users.
  3. Operators cannot reliably reverse a harmful release or identify what changed.
  4. Reported generalization may be overstated because the test set contains known information.

Explanation

Correct Answer

A:

For the risk comparison, quality can degrade gradually without generating a conspicuous technical outage. The shortcut overlooks the central safeguard: set post-release alerts for changes in input populations, outcomes and model quality. Real-world distributions can change after launch, making old validation evidence insufficient

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

C:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

D:

This consequence follows more directly from another shortcut: Randomly divide rows even when identical customers and future outcomes occur in both sets. The shortcut actually described would instead create this risk: Quality can degrade gradually without generating a conspicuous technical outage.

Question 530

Choose the intervention and validation that fit this business problem: A flood warning model begins receiving readings from a replacement sensor network.

  1. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  3. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.
  4. Action: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. Evidence: Review a documented impact assessment with risk owners, affected groups and recorded design responses.

Explanation

Correct Answer

B:

For the action and verification, set post-release alerts for changes in input populations, outcomes and model quality. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed. Together, they implement and test the intended safeguard. Real-world distributions can change after launch, making old validation evidence insufficient

Incorrect Answers

A:

This pair supports another governance intervention: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

C:

This pair supports another governance intervention: Version models, prompts and data transformations and define tested rollback criteria. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

D:

This pair supports another governance intervention: Evaluate plausible benefits, harms, likelihood and severity at the design stage while alternatives remain open. For the actual case, the required action is: Set post-release alerts for changes in input populations, outcomes and model quality. The supporting evidence is: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.

Question 531

A translation bot gives dangerous evacuation instructions and a team fixes only one phrase. Which governance action most directly addresses the need in this situation?

  1. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  2. Set post-release alerts for changes in input populations, outcomes and model quality.
  3. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.
  4. Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

Explanation

Correct Answer

C:

Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The practical acceptance evidence is: confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Incorrect Answers

A:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The more direct step is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

B:

This would help address a separate governance task: Set post-release alerts for changes in input populations, outcomes and model quality. For this case, however, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The more direct step is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

D:

This would help address a separate governance task: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For this case, however, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The more direct step is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

Question 532

A translation bot gives dangerous evacuation instructions and a team fixes only one phrase. Which explanation best connects the risk to the proposed governance action?

  1. Controls may erode as use cases, personnel, vendors and model dependencies change.
  2. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  3. A dataset can be large yet systematically miss important subgroups or operating conditions.
  4. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Explanation

Correct Answer

D:

A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. That is why the case calls for this action: document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

Incorrect Answers

A:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

B:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

C:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Question 533

The organization has implemented a control for this case: A translation bot gives dangerous evacuation instructions and a team fixes only one phrase. Which result would best validate it?

  1. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  2. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  3. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  4. Sample labeled records across sources and measure disagreement, missingness and remediation results.

Explanation

Correct Answer

A:

Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Incorrect Answers

B:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

C:

Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

D:

Sample labeled records across sources and measure disagreement, missingness and remediation results would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Question 534

For this case—A translation bot gives dangerous evacuation instructions and a team fixes only one phrase—staff choose the shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. Which consequence best explains why this is inadequate?

  1. Users and downstream operators may rely on outdated or misleading guidance.
  2. Recurring systemic causes may survive the immediate patch and harm additional users.
  3. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  4. The model can perform badly on neglected populations despite strong headline metrics.

Explanation

Correct Answer

B:

Recurring systemic causes may survive the immediate patch and harm additional users. The shortcut overlooks the central safeguard: document observed harm, affected scope, contributing technical and organizational factors, and corrective action. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Reuse old disclosure text after changing the model purpose and acceptable uses. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users.

C:

This consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users.

D:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users.

Question 535

Choose the intervention and validation that fit this business problem: A translation bot gives dangerous evacuation instructions and a team fixes only one phrase.

  1. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  2. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  4. Action: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. Evidence: Sample labeled records across sources and measure disagreement, missingness and remediation results.

Explanation

Correct Answer

C:

Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions. Together, they implement and test the intended safeguard. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Incorrect Answers

A:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

B:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

D:

This pair supports another governance intervention: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Question 536

An AI project team faces this situation: A scoring engine misclassifies a customer group after a data pipeline update and the alert is dismissed. What should the team do first to manage the central issue?

  1. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  2. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  3. Prevent leakage of future information or duplicate records across training and independent evaluation sets.
  4. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

Explanation

Correct Answer

D:

For the decision, document observed harm, affected scope, contributing technical and organizational factors, and corrective action. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The practical acceptance evidence is: confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions

Incorrect Answers

A:

This would help address a separate governance task: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For this case, however, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The more direct step is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

B:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—is not the most direct response to this case. The matching action is to document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

C:

This would help address a separate governance task: Prevent leakage of future information or duplicate records across training and independent evaluation sets. For this case, however, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. The more direct step is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.

Question 537

A scoring engine misclassifies a customer group after a data pipeline update and the alert is dismissed. Which explanation best connects the risk to the proposed governance action?

  1. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.
  2. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.
  3. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  4. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Explanation

Correct Answer

A:

For the technical explanation, a service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls. That is why the case calls for this action: document observed harm, affected scope, contributing technical and organizational factors, and corrective action

Incorrect Answers

B:

That explains why a team might run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

C:

For the technical explanation, that explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls

D:

That explains why a team might write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria, but it does not justify the decision required by the case. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.

Question 538

A scoring engine misclassifies a customer group after a data pipeline update and the alert is dismissed. Which recorded test or observation most directly demonstrates an effective response?

  1. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  4. Validate the use-case specification with business and affected-user representatives before model selection.

Explanation

Correct Answer

B:

For the verification, confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls

Incorrect Answers

A:

Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

C:

For the verification, compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions

D:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Question 539

A team considers this AI use: A scoring engine misclassifies a customer group after a data pipeline update and the alert is dismissed. Rather than applying the relevant control, it will close every harmful AI event as a generic software bug without investigating model behavior. What can go wrong?

  1. Previously approved controls and thresholds may no longer match the deployed model.
  2. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  3. Recurring systemic causes may survive the immediate patch and harm additional users.
  4. Known risks may remain uncontrolled despite formal notice to users.

Explanation

Correct Answer

C:

For the risk comparison, recurring systemic causes may survive the immediate patch and harm additional users. The shortcut overlooks the central safeguard: document observed harm, affected scope, contributing technical and organizational factors, and corrective action. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users

D:

This consequence follows more directly from another shortcut: Publish a warning label while retaining a design with an avoidable unsafe autonomous action. The shortcut actually described would instead create this risk: Recurring systemic causes may survive the immediate patch and harm additional users.

Question 540

The governance team needs both a decision and a way to verify it for this case: A scoring engine misclassifies a customer group after a data pipeline update and the alert is dismissed. Which pair is best?

  1. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  2. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  3. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  4. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Explanation

Correct Answer

D:

For the action and verification, document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions. Together, they implement and test the intended safeguard. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions

B:

This pair supports another governance intervention: Set post-release alerts for changes in input populations, outcomes and model quality. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

C:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. The supporting evidence is: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.

Question 541

A utility launches a new outage-priority model that immediately gives unstable rankings. Which governance action most directly addresses the need in this situation?

  1. Version models, prompts and data transformations and define tested rollback criteria.
  2. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  3. Document observed harm, affected scope, contributing technical and organizational factors, and corrective action.
  4. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Explanation

Correct Answer

A:

Version models, prompts and data transformations and define tested rollback criteria. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The practical acceptance evidence is: simulate rollback to a prior approved version and verify restored outputs and access policy.

Incorrect Answers

B:

This would help address a separate governance task: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

C:

This would help address a separate governance task: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

D:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

Question 542

A utility launches a new outage-priority model that immediately gives unstable rankings. Why is the recommended governance approach appropriate?

  1. Controls may erode as use cases, personnel, vendors and model dependencies change.
  2. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  3. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  4. Traceability supports review of how a model was chosen and why residual risks were considered acceptable.

Explanation

Correct Answer

B:

Controlled versioning enables recovery when a new system revision introduces unacceptable behavior. That is why the case calls for this action: version models, prompts and data transformations and define tested rollback criteria.

Incorrect Answers

A:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

C:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

D:

That explains why a team might maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Question 543

An AI governance reviewer examines this situation: A utility launches a new outage-priority model that immediately gives unstable rankings. Which acceptance check is most decisive?

  1. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  2. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Simulate rollback to a prior approved version and verify restored outputs and access policy.
  4. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Explanation

Correct Answer

C:

Simulate rollback to a prior approved version and verify restored outputs and access policy. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Incorrect Answers

A:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

B:

Review the adversarial test inventory, observed failures, remediations and retest outcomes would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

D:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

Question 544

For this case—A utility launches a new outage-priority model that immediately gives unstable rankings—staff choose the shortcut: Overwrite production model weights and prompts without retaining an approved prior version. Which consequence best explains why this is inadequate?

  1. Production begins without a defensible record of known limits or release acceptance.
  2. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  3. The model can learn labeling errors and reproduce inconsistent operational decisions.
  4. Operators cannot reliably reverse a harmful release or identify what changed.

Explanation

Correct Answer

D:

Operators cannot reliably reverse a harmful release or identify what changed. The shortcut overlooks the central safeguard: version models, prompts and data transformations and define tested rollback criteria. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed.

B:

This consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed.

C:

This consequence follows more directly from another shortcut: Treat every inherited annotation as correct because it came from a large historical database. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed.

Question 545

A utility launches a new outage-priority model that immediately gives unstable rankings. Which action and evidence pair most directly addresses the situation?

  1. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.
  2. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  3. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  4. Action: Include privacy, adversarial, access-control and integration tests in training and system evaluation. Evidence: Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.

Explanation

Correct Answer

A:

Version models, prompts and data transformations and define tested rollback criteria. Simulate rollback to a prior approved version and verify restored outputs and access policy. Together, they implement and test the intended safeguard. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Incorrect Answers

B:

This pair supports another governance intervention: Require notification, evaluation and approval paths for significant updates to externally supplied models. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

C:

This pair supports another governance intervention: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

D:

This pair supports another governance intervention: Include privacy, adversarial, access-control and integration tests in training and system evaluation. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

Question 546

A public information assistant changes its retrieval prompt and begins quoting irrelevant documents. Which step would provide the most targeted control for the problem described?

  1. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  2. Version models, prompts and data transformations and define tested rollback criteria.
  3. Set post-release alerts for changes in input populations, outcomes and model quality.
  4. Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria.

Explanation

Correct Answer

B:

For the decision, version models, prompts and data transformations and define tested rollback criteria. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The practical acceptance evidence is: simulate rollback to a prior approved version and verify restored outputs and access policy

Incorrect Answers

A:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

C:

This would help address a separate governance task: Set post-release alerts for changes in input populations, outcomes and model quality. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

D:

This would help address a separate governance task: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For this case, however, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. The more direct step is: Version models, prompts and data transformations and define tested rollback criteria.

Question 547

A team reviewing the following situation asks for the most relevant governance principle: A public information assistant changes its retrieval prompt and begins quoting irrelevant documents. Which principle best explains the required control?

  1. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.
  2. Transparency materials must reflect the actual system version and relevant role or legal requirements.
  3. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  4. An infrequent but catastrophic error may require stronger controls than a common minor inconvenience.

Explanation

Correct Answer

C:

For the technical explanation, controlled versioning enables recovery when a new system revision introduces unacceptable behavior. That is why the case calls for this action: version models, prompts and data transformations and define tested rollback criteria

Incorrect Answers

A:

That explains why a team might run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

B:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

D:

That explains why a team might assess AI failure risks using both estimated occurrence and potential magnitude of harm, but it does not justify the decision required by the case. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.

Question 548

A public information assistant changes its retrieval prompt and begins quoting irrelevant documents. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  3. Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios.
  4. Simulate rollback to a prior approved version and verify restored outputs and access policy.

Explanation

Correct Answer

D:

For the verification, simulate rollback to a prior approved version and verify restored outputs and access policy. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior

Incorrect Answers

A:

Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

B:

Compare deployed behavior and current instructions with approved disclosures and update records as needed would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

C:

Execute prompt-injection, privacy leakage and authorization tests using representative threat scenarios would test a different outcome. The directly relevant evidence is to simulate rollback to a prior approved version and verify restored outputs and access policy.

Question 549

For this case—A public information assistant changes its retrieval prompt and begins quoting irrelevant documents—staff choose the shortcut: Overwrite production model weights and prompts without retaining an approved prior version. Which consequence best explains why this is inadequate?

  1. Operators cannot reliably reverse a harmful release or identify what changed.
  2. Production begins without a defensible record of known limits or release acceptance.
  3. Prompt abuse, unsafe escalation and policy bypass may remain undetected.
  4. The solution may optimize an irrelevant metric while exposing people to unassessed harms.

Explanation

Correct Answer

A:

For the risk comparison, operators cannot reliably reverse a harmful release or identify what changed. The shortcut overlooks the central safeguard: version models, prompts and data transformations and define tested rollback criteria. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed

C:

For the risk comparison, this consequence follows more directly from another shortcut: Conclude an assistant is safe after ten ordinary successful user requests. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed

D:

This consequence follows more directly from another shortcut: Choose the highest-scoring model benchmark first and define the business objective afterward. The shortcut actually described would instead create this risk: Operators cannot reliably reverse a harmful release or identify what changed.

Question 550

A public information assistant changes its retrieval prompt and begins quoting irrelevant documents. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  2. Action: Version models, prompts and data transformations and define tested rollback criteria. Evidence: Simulate rollback to a prior approved version and verify restored outputs and access policy.
  3. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  4. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Explanation

Correct Answer

B:

For the action and verification, version models, prompts and data transformations and define tested rollback criteria. Simulate rollback to a prior approved version and verify restored outputs and access policy. Together, they implement and test the intended safeguard. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior

Incorrect Answers

A:

This pair supports another governance intervention: Set post-release alerts for changes in input populations, outcomes and model quality. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

C:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

D:

This pair supports another governance intervention: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For the actual case, the required action is: Version models, prompts and data transformations and define tested rollback criteria. The supporting evidence is: Simulate rollback to a prior approved version and verify restored outputs and access policy.

Question 551

A long-running aid eligibility system has never repeated subgroup fairness testing after policy changes. Which governance action most directly addresses the need in this situation?

  1. Set post-release alerts for changes in input populations, outcomes and model quality.
  2. Version models, prompts and data transformations and define tested rollback criteria.
  3. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  4. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.

Explanation

Correct Answer

C:

Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Controls may erode as use cases, personnel, vendors and model dependencies change. The practical acceptance evidence is: check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Incorrect Answers

A:

This would help address a separate governance task: Set post-release alerts for changes in input populations, outcomes and model quality. For this case, however, controls may erode as use cases, personnel, vendors and model dependencies change. The more direct step is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

B:

This would help address a separate governance task: Version models, prompts and data transformations and define tested rollback criteria. For this case, however, controls may erode as use cases, personnel, vendors and model dependencies change. The more direct step is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

D:

This would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, controls may erode as use cases, personnel, vendors and model dependencies change. The more direct step is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

Question 552

A long-running aid eligibility system has never repeated subgroup fairness testing after policy changes. Which explanation best connects the risk to the proposed governance action?

  1. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  2. A reproducible release package connects intended use and tested limitations to an accountable deployment decision.
  3. A dataset can be large yet systematically miss important subgroups or operating conditions.
  4. Controls may erode as use cases, personnel, vendors and model dependencies change.

Explanation

Correct Answer

D:

Controls may erode as use cases, personnel, vendors and model dependencies change. That is why the case calls for this action: plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

Incorrect Answers

A:

That explains why a team might require notification, evaluation and approval paths for significant updates to externally supplied models, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

B:

That explains why a team might require a model or system card, deployment limits, safety tests and responsible release approval before production, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

C:

That explains why a team might evaluate whether training and validation samples cover the populations and conditions the model will encounter, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

Question 553

A long-running aid eligibility system has never repeated subgroup fairness testing after policy changes. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Simulate rollback to a prior approved version and verify restored outputs and access policy.
  4. Validate the use-case specification with business and affected-user representatives before model selection.

Explanation

Correct Answer

A:

Check the assurance calendar, completed review evidence and tracking of overdue corrective actions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Controls may erode as use cases, personnel, vendors and model dependencies change.

Incorrect Answers

B:

Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

C:

Simulate rollback to a prior approved version and verify restored outputs and access policy would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

D:

Validate the use-case specification with business and affected-user representatives before model selection would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Question 554

The organization has this problem: A long-running aid eligibility system has never repeated subgroup fairness testing after policy changes. It takes a shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. What is the principal risk?

  1. Recurring systemic causes may survive the immediate patch and harm additional users.
  2. Subsequent drift or control failures can remain unnoticed.
  3. Quality can degrade gradually without generating a conspicuous technical outage.
  4. Auditors and incident responders cannot recover why critical choices were made.

Explanation

Correct Answer

B:

Subsequent drift or control failures can remain unnoticed. The shortcut overlooks the central safeguard: plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Controls may erode as use cases, personnel, vendors and model dependencies change.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed.

C:

This consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed.

D:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed.

Question 555

A long-running aid eligibility system has never repeated subgroup fairness testing after policy changes. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  4. Action: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. Evidence: Validate the use-case specification with business and affected-user representatives before model selection.

Explanation

Correct Answer

C:

Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions. Together, they implement and test the intended safeguard. Controls may erode as use cases, personnel, vendors and model dependencies change.

Incorrect Answers

A:

This pair supports another governance intervention: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

B:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

D:

This pair supports another governance intervention: Write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Question 556

An AI project team faces this situation: A transport operator keeps its AI camera model for years without retesting under new road conditions. What should the team do first to manage the central issue?

  1. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  2. Require a model or system card, deployment limits, safety tests and responsible release approval before production.
  3. Identify directly and indirectly affected groups and incorporate their concerns into design reviews.
  4. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

Explanation

Correct Answer

D:

For the decision, plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Controls may erode as use cases, personnel, vendors and model dependencies change. The practical acceptance evidence is: check the assurance calendar, completed review evidence and tracking of overdue corrective actions

Incorrect Answers

A:

This would help address a separate governance task: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For this case, however, controls may erode as use cases, personnel, vendors and model dependencies change. The more direct step is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

B:

The alternative—require a model or system card, deployment limits, safety tests and responsible release approval before production—is not the most direct response to this case. The matching action is to plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.

C:

For the decision, this would help address a separate governance task: Identify directly and indirectly affected groups and incorporate their concerns into design reviews. For this case, however, controls may erode as use cases, personnel, vendors and model dependencies change. The more direct step is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests

Question 557

A transport operator keeps its AI camera model for years without retesting under new road conditions. Why is the recommended governance approach appropriate?

  1. Controls may erode as use cases, personnel, vendors and model dependencies change.
  2. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  3. Real-world distributions can change after launch, making old validation evidence insufficient.
  4. Reliable provenance helps assess rights, representativeness and the source of discovered model failures.

Explanation

Correct Answer

A:

For the technical explanation, controls may erode as use cases, personnel, vendors and model dependencies change. That is why the case calls for this action: plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests

Incorrect Answers

B:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

C:

That explains why a team might set post-release alerts for changes in input populations, outcomes and model quality, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

D:

That explains why a team might record data origin, collection context, permitted use, transformations and lineage for training inputs, but it does not justify the decision required by the case. Controls may erode as use cases, personnel, vendors and model dependencies change.

Question 558

An AI governance reviewer examines this situation: A transport operator keeps its AI camera model for years without retesting under new road conditions. Which acceptance check is most decisive?

  1. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  2. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  3. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  4. Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.

Explanation

Correct Answer

B:

For the verification, check the assurance calendar, completed review evidence and tracking of overdue corrective actions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Controls may erode as use cases, personnel, vendors and model dependencies change

Incorrect Answers

A:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

C:

Compare deployed behavior and current instructions with approved disclosures and update records as needed would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

D:

Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals would test a different outcome. The directly relevant evidence is to check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Question 559

The organization has this problem: A transport operator keeps its AI camera model for years without retesting under new road conditions. It takes a shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. What is the principal risk?

  1. Users and downstream operators may rely on outdated or misleading guidance.
  2. Quality can degrade gradually without generating a conspicuous technical outage.
  3. Subsequent drift or control failures can remain unnoticed.
  4. The model can perform badly on neglected populations despite strong headline metrics.

Explanation

Correct Answer

C:

For the risk comparison, subsequent drift or control failures can remain unnoticed. The shortcut overlooks the central safeguard: plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Controls may erode as use cases, personnel, vendors and model dependencies change

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Reuse old disclosure text after changing the model purpose and acceptable uses. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed

D:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Subsequent drift or control failures can remain unnoticed.

Question 560

A transport operator keeps its AI camera model for years without retesting under new road conditions. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  2. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.
  4. Action: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Evidence: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Explanation

Correct Answer

D:

For the action and verification, plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions. Together, they implement and test the intended safeguard. Controls may erode as use cases, personnel, vendors and model dependencies change

Incorrect Answers

A:

This pair supports another governance intervention: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

B:

This pair supports another governance intervention: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

C:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests. The supporting evidence is: Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.

Question 561

Consider the following proposed AI use: A bank's third-party fraud API silently changes its scoring model during a peak season. Which response best addresses the governance gap?

  1. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  2. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  3. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  4. Prevent leakage of future information or duplicate records across training and independent evaluation sets.

Explanation

Correct Answer

A:

Require notification, evaluation and approval paths for significant updates to externally supplied models. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same. The practical acceptance evidence is: re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

Incorrect Answers

B:

The alternative—maintain accurate user instructions, technical limitations and any required public disclosures for the released model—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

C:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

D:

The alternative—prevent leakage of future information or duplicate records across training and independent evaluation sets—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Question 562

A bank's third-party fraud API silently changes its scoring model during a peak season. Which explanation best connects the risk to the proposed governance action?

  1. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  2. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  3. Transparency materials must reflect the actual system version and relevant role or legal requirements.
  4. Governance cannot evaluate fitness or risk without knowing what decisions the AI is meant to influence.

Explanation

Correct Answer

B:

A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same. That is why the case calls for this action: require notification, evaluation and approval paths for significant updates to externally supplied models.

Incorrect Answers

A:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

C:

That explains why a team might maintain accurate user instructions, technical limitations and any required public disclosures for the released model, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

D:

That explains why a team might write a bounded AI use-case statement with intended users, supported decisions and measurable success criteria, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Question 563

A bank's third-party fraud API silently changes its scoring model during a peak season. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Check the assurance calendar, completed review evidence and tracking of overdue corrective actions.
  2. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  3. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  4. Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Explanation

Correct Answer

C:

Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Incorrect Answers

A:

The alternative—check the assurance calendar, completed review evidence and tracking of overdue corrective actions—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

B:

The alternative—review the adversarial test inventory, observed failures, remediations and retest outcomes—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

D:

The alternative—observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Question 564

For this case—A bank's third-party fraud API silently changes its scoring model during a peak season—staff choose the shortcut: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. Which consequence best explains why this is inadequate?

  1. Users and downstream operators may rely on outdated or misleading guidance.
  2. Operators cannot reliably reverse a harmful release or identify what changed.
  3. The model can perform badly on neglected populations despite strong headline metrics.
  4. Previously approved controls and thresholds may no longer match the deployed model.

Explanation

Correct Answer

D:

Previously approved controls and thresholds may no longer match the deployed model. The shortcut overlooks the central safeguard: require notification, evaluation and approval paths for significant updates to externally supplied models. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Reuse old disclosure text after changing the model purpose and acceptable uses. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

B:

This consequence follows more directly from another shortcut: Overwrite production model weights and prompts without retaining an approved prior version. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

C:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

Question 565

The governance team needs both a decision and a way to verify it for this case: A bank's third-party fraud API silently changes its scoring model during a peak season. Which pair is best?

  1. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  2. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  3. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Action: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. Evidence: Observe representative cases to confirm reviewers can interpret evidence, challenge outputs and halt decisions.

Explanation

Correct Answer

A:

Require notification, evaluation and approval paths for significant updates to externally supplied models. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change. Together, they implement and test the intended safeguard. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Incorrect Answers

B:

This pair supports another governance intervention: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

C:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

D:

This pair supports another governance intervention: Assign a trained reviewer authority, decision information and a realistic opportunity to intervene. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

Question 566

Consider the following proposed AI use: A hospital assistant receives a new foundation-model version and starts providing different advice. Which response best addresses the governance gap?

  1. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  2. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  3. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  4. Audit missing values, inconsistent labels and measurement errors before trusting training outcomes.

Explanation

Correct Answer

B:

For the decision, require notification, evaluation and approval paths for significant updates to externally supplied models. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same. The practical acceptance evidence is: re-evaluate sampled system behavior and contractual obligations after a documented vendor model change

Incorrect Answers

A:

For the decision, the alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

C:

For the decision, the alternative—maintain accurate user instructions, technical limitations and any required public disclosures for the released model—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

D:

This would help address a separate governance task: Audit missing values, inconsistent labels and measurement errors before trusting training outcomes. For this case, however, a vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same. The more direct step is: Require notification, evaluation and approval paths for significant updates to externally supplied models.

Question 567

The organization chooses to require notification, evaluation and approval paths for significant updates to externally supplied models for this situation: A hospital assistant receives a new foundation-model version and starts providing different advice. Which rationale best supports that choice?

  1. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  2. Controls may erode as use cases, personnel, vendors and model dependencies change.
  3. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.
  4. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Explanation

Correct Answer

C:

For the technical explanation, a vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same. That is why the case calls for this action: require notification, evaluation and approval paths for significant updates to externally supplied models

Incorrect Answers

A:

For the technical explanation, that explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

B:

That explains why a team might plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

D:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Question 568

The organization has implemented a control for this case: A hospital assistant receives a new foundation-model version and starts providing different advice. Which result would best validate it?

  1. Simulate rollback to a prior approved version and verify restored outputs and access policy.
  2. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Trace selected training records back to licensed or approved origins through recorded preprocessing steps.
  4. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

Explanation

Correct Answer

D:

For the verification, re-evaluate sampled system behavior and contractual obligations after a documented vendor model change. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

Incorrect Answers

A:

The alternative—simulate rollback to a prior approved version and verify restored outputs and access policy—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

B:

The alternative—confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

C:

The alternative—trace selected training records back to licensed or approved origins through recorded preprocessing steps—would not resolve the stated issue. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same.

Question 569

A hospital assistant receives a new foundation-model version and starts providing different advice. The team instead proposes: Allow automatic supplier upgrades to high-impact AI workflows without regression testing. Which failure should the reviewer expect most directly?

  1. Previously approved controls and thresholds may no longer match the deployed model.
  2. Production begins without a defensible record of known limits or release acceptance.
  3. Quality can degrade gradually without generating a conspicuous technical outage.
  4. Auditors and incident responders cannot recover why critical choices were made.

Explanation

Correct Answer

A:

For the risk comparison, previously approved controls and thresholds may no longer match the deployed model. The shortcut overlooks the central safeguard: require notification, evaluation and approval paths for significant updates to externally supplied models. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

C:

This consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

D:

This consequence follows more directly from another shortcut: Keep only the final model file and discard the records of rejected designs and control decisions. The shortcut actually described would instead create this risk: Previously approved controls and thresholds may no longer match the deployed model.

Question 570

A hospital assistant receives a new foundation-model version and starts providing different advice. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  2. Action: Require notification, evaluation and approval paths for significant updates to externally supplied models. Evidence: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.
  3. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Action: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. Evidence: Review stratified metrics, stress cases and documented thresholds with accountable reviewers.

Explanation

Correct Answer

B:

For the action and verification, require notification, evaluation and approval paths for significant updates to externally supplied models. Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change. Together, they implement and test the intended safeguard. A vendor update can alter accuracy, safety, data processing or behavior even if an API endpoint remains the same

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change

C:

For the action and verification, this pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change

D:

This pair supports another governance intervention: Test predictive performance and failure modes across relevant subgroups, edge cases and realistic perturbations. For the actual case, the required action is: Require notification, evaluation and approval paths for significant updates to externally supplied models. The supporting evidence is: Re-evaluate sampled system behavior and contractual obligations after a documented vendor model change.

Question 571

Consider the following proposed AI use: A vendor extends a clinical decision-support model to a new population but leaves its instructions unchanged. Which response best addresses the governance gap?

  1. Plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests.
  2. Require notification, evaluation and approval paths for significant updates to externally supplied models.
  3. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.
  4. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.

Explanation

Correct Answer

C:

Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Transparency materials must reflect the actual system version and relevant role or legal requirements. The practical acceptance evidence is: compare deployed behavior and current instructions with approved disclosures and update records as needed.

Incorrect Answers

A:

The alternative—plan recurring assessments of model performance, fairness, security and governance controls rather than relying only on launch tests—would not resolve the stated issue. Transparency materials must reflect the actual system version and relevant role or legal requirements.

B:

This would help address a separate governance task: Require notification, evaluation and approval paths for significant updates to externally supplied models. For this case, however, transparency materials must reflect the actual system version and relevant role or legal requirements. The more direct step is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

D:

This would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, transparency materials must reflect the actual system version and relevant role or legal requirements. The more direct step is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

Question 572

The organization chooses to maintain accurate user instructions, technical limitations and any required public disclosures for the released model for this situation: A vendor extends a clinical decision-support model to a new population but leaves its instructions unchanged. Which rationale best supports that choice?

  1. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.
  2. Friendly demonstration prompts cannot reveal all unsafe behaviors or malicious interaction paths.
  3. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.
  4. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Explanation

Correct Answer

D:

Transparency materials must reflect the actual system version and relevant role or legal requirements. That is why the case calls for this action: maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

Incorrect Answers

A:

That explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements.

B:

That explains why a team might run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements.

C:

That explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Question 573

A vendor extends a clinical decision-support model to a new population but leaves its instructions unchanged. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  2. Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  3. Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  4. Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards.

Explanation

Correct Answer

A:

Compare deployed behavior and current instructions with approved disclosures and update records as needed. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Incorrect Answers

B:

Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

C:

Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

D:

Verify a risk register records severity, likelihood, assumptions and treatment decisions for key hazards would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

Question 574

For this case—A vendor extends a clinical decision-support model to a new population but leaves its instructions unchanged—staff choose the shortcut: Reuse old disclosure text after changing the model purpose and acceptable uses. Which consequence best explains why this is inadequate?

  1. Recurring systemic causes may survive the immediate patch and harm additional users.
  2. Users and downstream operators may rely on outdated or misleading guidance.
  3. Subsequent drift or control failures can remain unnoticed.
  4. Reported generalization may be overstated because the test set contains known information.

Explanation

Correct Answer

B:

Users and downstream operators may rely on outdated or misleading guidance. The shortcut overlooks the central safeguard: maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Close every harmful AI event as a generic software bug without investigating model behavior. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

C:

This consequence follows more directly from another shortcut: Treat launch-day testing as sufficient assurance for the model's entire operating life. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

D:

This consequence follows more directly from another shortcut: Randomly divide rows even when identical customers and future outcomes occur in both sets. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

Question 575

A vendor extends a clinical decision-support model to a new population but leaves its instructions unchanged. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Set post-release alerts for changes in input populations, outcomes and model quality. Evidence: Compare live input and outcome distributions with baselines and trigger reviews when thresholds are crossed.
  2. Action: Require a model or system card, deployment limits, safety tests and responsible release approval before production. Evidence: Inspect the signed release dossier for version, intended use, risk disposition, test coverage and rollback owner.
  3. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  4. Action: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. Evidence: Trace high-risk findings to elimination, reduction, controls and documented residual-risk acceptance.

Explanation

Correct Answer

C:

Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Compare deployed behavior and current instructions with approved disclosures and update records as needed. Together, they implement and test the intended safeguard. Transparency materials must reflect the actual system version and relevant role or legal requirements.

Incorrect Answers

A:

This pair supports another governance intervention: Set post-release alerts for changes in input populations, outcomes and model quality. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

B:

This pair supports another governance intervention: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

D:

This pair supports another governance intervention: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Question 576

An AI loan-scoring supplier updates its features without refreshing customer-facing limitation notices. Which governance action most directly addresses the need in this situation?

  1. Require a model or system card, deployment limits, safety tests and responsible release approval before production.
  2. Set post-release alerts for changes in input populations, outcomes and model quality.
  3. First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings.
  4. Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

Explanation

Correct Answer

D:

For the decision, maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Transparency materials must reflect the actual system version and relevant role or legal requirements. The practical acceptance evidence is: compare deployed behavior and current instructions with approved disclosures and update records as needed

Incorrect Answers

A:

This would help address a separate governance task: Require a model or system card, deployment limits, safety tests and responsible release approval before production. For this case, however, transparency materials must reflect the actual system version and relevant role or legal requirements. The more direct step is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

B:

This would help address a separate governance task: Set post-release alerts for changes in input populations, outcomes and model quality. For this case, however, transparency materials must reflect the actual system version and relevant role or legal requirements. The more direct step is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model.

C:

For the decision, this would help address a separate governance task: First consider eliminating harmful AI uses or redesigning risky features before relying solely on warnings. For this case, however, transparency materials must reflect the actual system version and relevant role or legal requirements. The more direct step is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model

Question 577

The organization chooses to maintain accurate user instructions, technical limitations and any required public disclosures for the released model for this situation: An AI loan-scoring supplier updates its features without refreshing customer-facing limitation notices. Which rationale best supports that choice?

  1. Transparency materials must reflect the actual system version and relevant role or legal requirements.
  2. Controlled versioning enables recovery when a new system revision introduces unacceptable behavior.
  3. A service outage label alone may miss data drift, unsafe prompts, missing review or inadequate controls.
  4. The system's users, decision subjects and impacted bystanders can experience different benefits and harms.

Explanation

Correct Answer

A:

For the technical explanation, transparency materials must reflect the actual system version and relevant role or legal requirements. That is why the case calls for this action: maintain accurate user instructions, technical limitations and any required public disclosures for the released model

Incorrect Answers

B:

That explains why a team might version models, prompts and data transformations and define tested rollback criteria, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements.

C:

For the technical explanation, that explains why a team might document observed harm, affected scope, contributing technical and organizational factors, and corrective action, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements

D:

For the technical explanation, that explains why a team might identify directly and indirectly affected groups and incorporate their concerns into design reviews, but it does not justify the decision required by the case. Transparency materials must reflect the actual system version and relevant role or legal requirements

Question 578

The organization has implemented a control for this case: An AI loan-scoring supplier updates its features without refreshing customer-facing limitation notices. Which result would best validate it?

  1. Simulate rollback to a prior approved version and verify restored outputs and access policy.
  2. Compare deployed behavior and current instructions with approved disclosures and update records as needed.
  3. Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  4. Check stakeholder maps and feedback records for representation of affected groups and mitigations.

Explanation

Correct Answer

B:

For the verification, compare deployed behavior and current instructions with approved disclosures and update records as needed. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Transparency materials must reflect the actual system version and relevant role or legal requirements

Incorrect Answers

A:

Simulate rollback to a prior approved version and verify restored outputs and access policy would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

C:

Review the adversarial test inventory, observed failures, remediations and retest outcomes would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

D:

Check stakeholder maps and feedback records for representation of affected groups and mitigations would test a different outcome. The directly relevant evidence is to compare deployed behavior and current instructions with approved disclosures and update records as needed.

Question 579

For this case—An AI loan-scoring supplier updates its features without refreshing customer-facing limitation notices—staff choose the shortcut: Reuse old disclosure text after changing the model purpose and acceptable uses. Which consequence best explains why this is inadequate?

  1. Quality can degrade gradually without generating a conspicuous technical outage.
  2. Production begins without a defensible record of known limits or release acceptance.
  3. Users and downstream operators may rely on outdated or misleading guidance.
  4. The model can perform badly on neglected populations despite strong headline metrics.

Explanation

Correct Answer

C:

For the risk comparison, users and downstream operators may rely on outdated or misleading guidance. The shortcut overlooks the central safeguard: maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Transparency materials must reflect the actual system version and relevant role or legal requirements

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume a model will remain accurate indefinitely because the training dataset was extensive. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

B:

This consequence follows more directly from another shortcut: Promote a model based only on a positive stakeholder demo with no versioned evidence. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

D:

This consequence follows more directly from another shortcut: Assume millions of data rows guarantee fairness and coverage for every user. The shortcut actually described would instead create this risk: Users and downstream operators may rely on outdated or misleading guidance.

Question 580

An AI loan-scoring supplier updates its features without refreshing customer-facing limitation notices. Which action and evidence pair most directly addresses the situation?

  1. Action: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. Evidence: Review the adversarial test inventory, observed failures, remediations and retest outcomes.
  2. Action: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. Evidence: Confirm an incident record links timeline, root-cause hypotheses, evidence, impact and retest decisions.
  3. Action: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. Evidence: Reconstruct a sample design decision from versioned requirements, alternatives, test results and approvals.
  4. Action: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Evidence: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Explanation

Correct Answer

D:

For the action and verification, maintain accurate user instructions, technical limitations and any required public disclosures for the released model. Compare deployed behavior and current instructions with approved disclosures and update records as needed. Together, they implement and test the intended safeguard. Transparency materials must reflect the actual system version and relevant role or legal requirements

Incorrect Answers

A:

This pair supports another governance intervention: Run targeted red-team and abuse tests against foreseeable misuse and boundary failures before the launch gate. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

B:

This pair supports another governance intervention: Document observed harm, affected scope, contributing technical and organizational factors, and corrective action. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

C:

This pair supports another governance intervention: Maintain dated design choices, model alternatives, risk decisions and stakeholder approvals through development. For the actual case, the required action is: Maintain accurate user instructions, technical limitations and any required public disclosures for the released model. The supporting evidence is: Compare deployed behavior and current instructions with approved disclosures and update records as needed.

Question 581

An AI project team faces this situation: A utility must predict equipment failure probabilities rather than draft an open-ended narrative. What should the team do first to manage the central issue?

  1. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  2. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  3. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  4. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

Explanation

Correct Answer

A:

Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Natural-language generation and structured classification solve different problems and need different evaluations. The practical acceptance evidence is: run representative task trials measuring utility, error types, explainability and required operating controls.

Incorrect Answers

B:

This would help address a separate governance task: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

C:

This would help address a separate governance task: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

D:

This would help address a separate governance task: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Question 582

A utility must predict equipment failure probabilities rather than draft an open-ended narrative. Why is the recommended governance approach appropriate?

  1. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  2. Natural-language generation and structured classification solve different problems and need different evaluations.
  3. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  4. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Explanation

Correct Answer

B:

Natural-language generation and structured classification solve different problems and need different evaluations. That is why the case calls for this action: compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Incorrect Answers

A:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations.

C:

That explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations.

D:

That explains why a team might define evidence access, incident notification and audit cooperation in third-party AI agreements, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations.

Question 583

A utility must predict equipment failure probabilities rather than draft an open-ended narrative. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  2. Observe simulated decisions with trained users and record comprehension, override and escalation results.
  3. Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Explanation

Correct Answer

C:

Run representative task trials measuring utility, error types, explainability and required operating controls. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Natural-language generation and structured classification solve different problems and need different evaluations.

Incorrect Answers

A:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls.

B:

Observe simulated decisions with trained users and record comprehension, override and escalation results would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls.

D:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls.

Question 584

A utility must predict equipment failure probabilities rather than draft an open-ended narrative. The team instead proposes: Adopt the largest generative model for every task without testing simpler alternatives. Which failure should the reviewer expect most directly?

  1. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  2. An agent may make unapproved purchases, record changes or external communications.
  3. Site-specific harms and legal obligations may remain unrecognized.
  4. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Explanation

Correct Answer

D:

Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice. The shortcut overlooks the central safeguard: compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Natural-language generation and structured classification solve different problems and need different evaluations.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

B:

This consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

C:

This consequence follows more directly from another shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Question 585

Choose the intervention and validation that fit this business problem: A utility must predict equipment failure probabilities rather than draft an open-ended narrative.

  1. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  2. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  4. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.

Explanation

Correct Answer

A:

Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Run representative task trials measuring utility, error types, explainability and required operating controls. Together, they implement and test the intended safeguard. Natural-language generation and structured classification solve different problems and need different evaluations.

Incorrect Answers

B:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls.

C:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls.

D:

This pair supports another governance intervention: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls.

Question 586

Consider the following proposed AI use: A public library wants to answer varied research questions and compares a classifier with a generative assistant. Which response best addresses the governance gap?

  1. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  2. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  3. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  4. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Explanation

Correct Answer

B:

For the decision, compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Natural-language generation and structured classification solve different problems and need different evaluations. The practical acceptance evidence is: run representative task trials measuring utility, error types, explainability and required operating controls

Incorrect Answers

A:

This would help address a separate governance task: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

C:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

D:

This would help address a separate governance task: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For this case, however, natural-language generation and structured classification solve different problems and need different evaluations. The more direct step is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Question 587

The organization chooses to compare classic predictive models with generative approaches against the defined problem and acceptable failure modes for this situation: A public library wants to answer varied research questions and compares a classifier with a generative assistant. Which rationale best supports that choice?

  1. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.
  2. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  3. Natural-language generation and structured classification solve different problems and need different evaluations.
  4. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Explanation

Correct Answer

C:

For the technical explanation, natural-language generation and structured classification solve different problems and need different evaluations. That is why the case calls for this action: compare classic predictive models with generative approaches against the defined problem and acceptable failure modes

Incorrect Answers

A:

That explains why a team might assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations.

B:

For the technical explanation, that explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations

D:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Natural-language generation and structured classification solve different problems and need different evaluations.

Question 588

A public library wants to answer varied research questions and compares a classifier with a generative assistant. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  3. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  4. Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

D:

For the verification, run representative task trials measuring utility, error types, explainability and required operating controls. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Natural-language generation and structured classification solve different problems and need different evaluations

Incorrect Answers

A:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls.

B:

For the verification, test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls

C:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to run representative task trials measuring utility, error types, explainability and required operating controls.

Question 589

For this case—A public library wants to answer varied research questions and compares a classifier with a generative assistant—staff choose the shortcut: Adopt the largest generative model for every task without testing simpler alternatives. Which consequence best explains why this is inadequate?

  1. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.
  2. An agent may make unapproved purchases, record changes or external communications.
  3. Answers can become stale and lack traceable sources after the source policy changes.
  4. Stakeholders may be misled and unable to take appropriate protective steps.

Explanation

Correct Answer

A:

For the risk comparison, unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice. The shortcut overlooks the central safeguard: compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Natural-language generation and structured classification solve different problems and need different evaluations

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice

C:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

D:

This consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Question 590

A public library wants to answer varied research questions and compares a classifier with a generative assistant. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  3. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  4. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Explanation

Correct Answer

B:

For the action and verification, compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Run representative task trials measuring utility, error types, explainability and required operating controls. Together, they implement and test the intended safeguard. Natural-language generation and structured classification solve different problems and need different evaluations

Incorrect Answers

A:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls.

C:

For the action and verification, this pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls

D:

This pair supports another governance intervention: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For the actual case, the required action is: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. The supporting evidence is: Run representative task trials measuring utility, error types, explainability and required operating controls.

Question 591

An AI project team faces this situation: A remote wildfire sensor must identify hazards even when internet access drops. What should the team do first to manage the central issue?

  1. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  2. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  3. Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.
  4. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

Explanation

Correct Answer

C:

Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The practical acceptance evidence is: document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Incorrect Answers

A:

This would help address a separate governance task: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

B:

This would help address a separate governance task: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

D:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

Question 592

A remote wildfire sensor must identify hazards even when internet access drops. Which explanation best connects the risk to the proposed governance action?

  1. Natural-language generation and structured classification solve different problems and need different evaluations.
  2. A larger model is not automatically the best option when the required task is narrow and resources are constrained.
  3. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.
  4. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Explanation

Correct Answer

D:

Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. That is why the case calls for this action: assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

Incorrect Answers

A:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

B:

That explains why a team might compare smaller and larger models on task performance, inference cost, energy use and maintainability, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

C:

That explains why a team might maintain usable override and escalation mechanisms for consequential AI-assisted decisions, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Question 593

A remote wildfire sensor must identify hazards even when internet access drops. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

A:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Incorrect Answers

B:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

C:

Run representative task trials measuring utility, error types, explainability and required operating controls would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

D:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Question 594

The organization has this problem: A remote wildfire sensor must identify hazards even when internet access drops. It takes a shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. What is the principal risk?

  1. Answers can become stale and lack traceable sources after the source policy changes.
  2. Connectivity and data-handling requirements may make a technically capable model unusable.
  3. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  4. Staff can act on fabricated answers or expose sensitive information through normal workflows.

Explanation

Correct Answer

B:

Connectivity and data-handling requirements may make a technically capable model unusable. The shortcut overlooks the central safeguard: assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable.

C:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable.

D:

This consequence follows more directly from another shortcut: Treat access to an AI tool as proof that employees know how to use it safely. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable.

Question 595

Choose the intervention and validation that fit this business problem: A remote wildfire sensor must identify hazards even when internet access drops.

  1. Action: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Evidence: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  2. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

C:

Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Document a fit assessment with real latency, connectivity, privacy and maintenance tests. Together, they implement and test the intended safeguard. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Incorrect Answers

A:

This pair supports another governance intervention: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

B:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

D:

This pair supports another governance intervention: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Question 596

A regulated laboratory considers cloud inference on restricted specimens without reviewing data transfer controls. Which governance action most directly addresses the need in this situation?

  1. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  2. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  3. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  4. Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

Explanation

Correct Answer

D:

For the decision, assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The practical acceptance evidence is: document a fit assessment with real latency, connectivity, privacy and maintenance tests

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints

B:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

C:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. The more direct step is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.

Question 597

A regulated laboratory considers cloud inference on restricted specimens without reviewing data transfer controls. Why is the recommended governance approach appropriate?

  1. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.
  2. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.
  3. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  4. Governance needs a response path for unsafe outcomes as well as conventional outages.

Explanation

Correct Answer

A:

For the technical explanation, hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations. That is why the case calls for this action: assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints

Incorrect Answers

B:

That explains why a team might specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

C:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

D:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Question 598

An AI governance reviewer examines this situation: A regulated laboratory considers cloud inference on restricted specimens without reviewing data transfer controls. Which acceptance check is most decisive?

  1. Observe simulated decisions with trained users and record comprehension, override and escalation results.
  2. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  3. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  4. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Explanation

Correct Answer

B:

For the verification, document a fit assessment with real latency, connectivity, privacy and maintenance tests. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations

Incorrect Answers

A:

Observe simulated decisions with trained users and record comprehension, override and escalation results would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

C:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

D:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Question 599

A team considers this AI use: A regulated laboratory considers cloud inference on restricted specimens without reviewing data transfer controls. Rather than applying the relevant control, it will send every real-time sensitive inference to a remote cloud service without examining operating conditions. What can go wrong?

  1. Cost and latency can rise without meaningful gains for the actual workflow.
  2. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  3. Connectivity and data-handling requirements may make a technically capable model unusable.
  4. A user can receive restricted information or trigger an unauthorized operation.

Explanation

Correct Answer

C:

For the risk comparison, connectivity and data-handling requirements may make a technically capable model unusable. The shortcut overlooks the central safeguard: assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Select a massive general-purpose model solely because its parameter count is highest. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable

D:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Connectivity and data-handling requirements may make a technically capable model unusable.

Question 600

Choose the intervention and validation that fit this business problem: A regulated laboratory considers cloud inference on restricted specimens without reviewing data transfer controls.

  1. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  2. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  3. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  4. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Explanation

Correct Answer

D:

For the action and verification, assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Document a fit assessment with real latency, connectivity, privacy and maintenance tests. Together, they implement and test the intended safeguard. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations

Incorrect Answers

A:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

B:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

C:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. The supporting evidence is: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Question 601

A university policy assistant must answer from frequently updated course rules. Which governance action most directly addresses the need in this situation?

  1. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  2. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  3. Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.
  4. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Explanation

Correct Answer

A:

Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs. The practical acceptance evidence is: compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Incorrect Answers

B:

The alternative—compare smaller and larger models on task performance, inference cost, energy use and maintainability—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

C:

The alternative—assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

D:

The alternative—implement and test a way to suspend unsafe AI functionality or switch to a safe fallback—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Question 602

The organization chooses to test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit for this situation: A university policy assistant must answer from frequently updated course rules. Which rationale best supports that choice?

  1. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  2. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  3. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.
  4. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Explanation

Correct Answer

B:

Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs. That is why the case calls for this action: test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.

Incorrect Answers

A:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

C:

That explains why a team might specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

D:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Question 603

A university policy assistant must answer from frequently updated course rules. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  3. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  4. Test users with different roles against protected documents and disallowed tool operations.

Explanation

Correct Answer

C:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Incorrect Answers

A:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

B:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

D:

Test users with different roles against protected documents and disallowed tool operations would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Question 604

The organization has this problem: A university policy assistant must answer from frequently updated course rules. It takes a shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. What is the principal risk?

  1. An agent may make unapproved purchases, record changes or external communications.
  2. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  3. Model harms and wrong decisions can rise while infrastructure remains healthy.
  4. Answers can become stale and lack traceable sources after the source policy changes.

Explanation

Correct Answer

D:

Answers can become stale and lack traceable sources after the source policy changes. The shortcut overlooks the central safeguard: test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes.

B:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes.

C:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes.

Question 605

A university policy assistant must answer from frequently updated course rules. Which action and evidence pair most directly addresses the situation?

  1. Action: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Evidence: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  2. Action: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Evidence: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  3. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Explanation

Correct Answer

A:

Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches. Together, they implement and test the intended safeguard. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Incorrect Answers

B:

This pair supports another governance intervention: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

C:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

D:

This pair supports another governance intervention: Monitor quality, safety, latency and feedback against meaningful production thresholds. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Question 606

A transport authority needs a service chatbot to cite the latest maintenance notices. Which step would provide the most targeted control for the problem described?

  1. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  2. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  3. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  4. Use region- and role-aware controls when the permitted AI use differs across operating locations.

Explanation

Correct Answer

B:

For the decision, test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs. The practical acceptance evidence is: compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches

Incorrect Answers

A:

The alternative—evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

C:

The alternative—compare classic predictive models with generative approaches against the defined problem and acceptable failure modes—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

D:

The alternative—use region- and role-aware controls when the permitted AI use differs across operating locations—would not resolve the stated issue. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Question 607

A team reviewing the following situation asks for the most relevant governance principle: A transport authority needs a service chatbot to cite the latest maintenance notices. Which principle best explains the required control?

  1. Natural-language generation and structured classification solve different problems and need different evaluations.
  2. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.
  3. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  4. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Explanation

Correct Answer

C:

For the technical explanation, retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs. That is why the case calls for this action: test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit

Incorrect Answers

A:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

B:

For the technical explanation, that explains why a team might specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs

D:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Question 608

A transport authority needs a service chatbot to cite the latest maintenance notices. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Audit overridden cases for decision context, reviewer authority and follow-through.
  4. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Explanation

Correct Answer

D:

For the verification, compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs

Incorrect Answers

A:

For the verification, document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches

B:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

C:

Audit overridden cases for decision context, reviewer authority and follow-through would test a different outcome. The directly relevant evidence is to compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Question 609

For this case—A transport authority needs a service chatbot to cite the latest maintenance notices—staff choose the shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. Which consequence best explains why this is inadequate?

  1. Answers can become stale and lack traceable sources after the source policy changes.
  2. An agent may make unapproved purchases, record changes or external communications.
  3. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  4. The system may fail on the very cases the deploying organization must handle.

Explanation

Correct Answer

A:

For the risk comparison, answers can become stale and lack traceable sources after the source policy changes. The shortcut overlooks the central safeguard: test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes

C:

For the risk comparison, this consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes

D:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: Answers can become stale and lack traceable sources after the source policy changes.

Question 610

The governance team needs both a decision and a way to verify it for this case: A transport authority needs a service chatbot to cite the latest maintenance notices. Which pair is best?

  1. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  2. Action: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Evidence: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  3. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Explanation

Correct Answer

B:

For the action and verification, test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches. Together, they implement and test the intended safeguard. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs

Incorrect Answers

A:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

C:

For the action and verification, this pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches

D:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. The supporting evidence is: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Question 611

An AI project team faces this situation: A procurement agent is proposed to negotiate and place orders without purchase limits. What should the team do first to manage the central issue?

  1. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  2. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  3. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  4. Use region- and role-aware controls when the permitted AI use differs across operating locations.

Explanation

Correct Answer

C:

Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act. The practical acceptance evidence is: test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Incorrect Answers

A:

The alternative—compare smaller and larger models on task performance, inference cost, energy use and maintainability—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

B:

The alternative—evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

D:

The alternative—use region- and role-aware controls when the permitted AI use differs across operating locations—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Question 612

A team reviewing the following situation asks for the most relevant governance principle: A procurement agent is proposed to negotiate and place orders without purchase limits. Which principle best explains the required control?

  1. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  2. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  3. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  4. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Explanation

Correct Answer

D:

Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act. That is why the case calls for this action: specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.

Incorrect Answers

A:

That explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

B:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

C:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Question 613

A procurement agent is proposed to negotiate and place orders without purchase limits. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  2. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  4. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Explanation

Correct Answer

A:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Incorrect Answers

B:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

C:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

D:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Question 614

For this case—A procurement agent is proposed to negotiate and place orders without purchase limits—staff choose the shortcut: Give an autonomous assistant broad write access because it handled informational questions well. Which consequence best explains why this is inadequate?

  1. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.
  2. An agent may make unapproved purchases, record changes or external communications.
  3. Connectivity and data-handling requirements may make a technically capable model unusable.
  4. Model harms and wrong decisions can rise while infrastructure remains healthy.

Explanation

Correct Answer

B:

An agent may make unapproved purchases, record changes or external communications. The shortcut overlooks the central safeguard: specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications.

C:

This consequence follows more directly from another shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications.

D:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications.

Question 615

Choose the intervention and validation that fit this business problem: A procurement agent is proposed to negotiate and place orders without purchase limits.

  1. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  2. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Action: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Evidence: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  4. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Explanation

Correct Answer

C:

Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks. Together, they implement and test the intended safeguard. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

B:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

D:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Question 616

A charity wants a donor assistant to draft responses and automatically change banking details. Which step would provide the most targeted control for the problem described?

  1. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  2. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  3. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  4. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.

Explanation

Correct Answer

D:

For the decision, specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act. The practical acceptance evidence is: test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks

Incorrect Answers

A:

The alternative—test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

B:

The alternative—compare classic predictive models with generative approaches against the defined problem and acceptable failure modes—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

C:

The alternative—verify model, dataset and output licensing conditions before integrating or redistributing AI components—would not resolve the stated issue. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Question 617

A charity wants a donor assistant to draft responses and automatically change banking details. Which explanation best connects the risk to the proposed governance action?

  1. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.
  2. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  3. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  4. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Explanation

Correct Answer

A:

For the technical explanation, tool-using systems can create side effects, so a plausible language response is not the same as authorization to act. That is why the case calls for this action: specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy

Incorrect Answers

B:

For the technical explanation, that explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act

C:

For the technical explanation, that explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act

D:

That explains why a team might assess actual affected users, decisions, data and harms in the intended deployment environment, but it does not justify the decision required by the case. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Question 618

A charity wants a donor assistant to draft responses and automatically change banking details. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  2. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  3. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

B:

For the verification, test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act

Incorrect Answers

A:

For the verification, compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks

C:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

D:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Question 619

A charity wants a donor assistant to draft responses and automatically change banking details. The team instead proposes: Give an autonomous assistant broad write access because it handled informational questions well. Which failure should the reviewer expect most directly?

  1. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  2. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.
  3. An agent may make unapproved purchases, record changes or external communications.
  4. Model harms and wrong decisions can rise while infrastructure remains healthy.

Explanation

Correct Answer

C:

For the risk comparison, an agent may make unapproved purchases, record changes or external communications. The shortcut overlooks the central safeguard: specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications

D:

For the risk comparison, this consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: An agent may make unapproved purchases, record changes or external communications

Question 620

The governance team needs both a decision and a way to verify it for this case: A charity wants a donor assistant to draft responses and automatically change banking details. Which pair is best?

  1. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  3. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  4. Action: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Evidence: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Explanation

Correct Answer

D:

For the action and verification, specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks. Together, they implement and test the intended safeguard. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act

Incorrect Answers

A:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

B:

For the action and verification, this pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks

C:

This pair supports another governance intervention: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For the actual case, the required action is: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. The supporting evidence is: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Question 621

A nonprofit needs to categorize short help tickets on a low-cost local appliance. Which governance action most directly addresses the need in this situation?

  1. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  2. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  3. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  4. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Explanation

Correct Answer

A:

Compare smaller and larger models on task performance, inference cost, energy use and maintainability. A larger model is not automatically the best option when the required task is narrow and resources are constrained. The practical acceptance evidence is: evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Incorrect Answers

B:

The alternative—specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy—is not the most direct response to this case. The matching action is to compare smaller and larger models on task performance, inference cost, energy use and maintainability.

C:

The alternative—test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit—is not the most direct response to this case. The matching action is to compare smaller and larger models on task performance, inference cost, energy use and maintainability.

D:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, a larger model is not automatically the best option when the required task is narrow and resources are constrained. The more direct step is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability.

Question 622

The organization chooses to compare smaller and larger models on task performance, inference cost, energy use and maintainability for this situation: A nonprofit needs to categorize short help tickets on a low-cost local appliance. Which rationale best supports that choice?

  1. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  2. A larger model is not automatically the best option when the required task is narrow and resources are constrained.
  3. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  4. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Explanation

Correct Answer

B:

A larger model is not automatically the best option when the required task is narrow and resources are constrained. That is why the case calls for this action: compare smaller and larger models on task performance, inference cost, energy use and maintainability.

Incorrect Answers

A:

That explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

C:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

D:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Question 623

An AI governance reviewer examines this situation: A nonprofit needs to categorize short help tickets on a low-cost local appliance. Which acceptance check is most decisive?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  3. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  4. Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

C:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Incorrect Answers

A:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

B:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

D:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Question 624

A team considers this AI use: A nonprofit needs to categorize short help tickets on a low-cost local appliance. Rather than applying the relevant control, it will select a massive general-purpose model solely because its parameter count is highest. What can go wrong?

  1. Connectivity and data-handling requirements may make a technically capable model unusable.
  2. An agent may make unapproved purchases, record changes or external communications.
  3. Unsafe behavior may continue without a responsible containment decision.
  4. Cost and latency can rise without meaningful gains for the actual workflow.

Explanation

Correct Answer

D:

Cost and latency can rise without meaningful gains for the actual workflow. The shortcut overlooks the central safeguard: compare smaller and larger models on task performance, inference cost, energy use and maintainability. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

B:

This consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

C:

This consequence follows more directly from another shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

Question 625

Choose the intervention and validation that fit this business problem: A nonprofit needs to categorize short help tickets on a low-cost local appliance.

  1. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  2. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  3. Action: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. Evidence: Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.
  4. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Explanation

Correct Answer

A:

Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements. Together, they implement and test the intended safeguard. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Incorrect Answers

B:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

C:

This pair supports another governance intervention: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

D:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Question 626

An AI project team faces this situation: A factory deploys visual defect detection at high speed on a constrained edge device. What should the team do first to manage the central issue?

  1. Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.
  2. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  3. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.
  4. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Explanation

Correct Answer

B:

For the decision, compare smaller and larger models on task performance, inference cost, energy use and maintainability. A larger model is not automatically the best option when the required task is narrow and resources are constrained. The practical acceptance evidence is: evaluate equivalent examples on accuracy, latency, operating expense and security requirements

Incorrect Answers

A:

The alternative—assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints—is not the most direct response to this case. The matching action is to compare smaller and larger models on task performance, inference cost, energy use and maintainability.

C:

The alternative—compare classic predictive models with generative approaches against the defined problem and acceptable failure modes—is not the most direct response to this case. The matching action is to compare smaller and larger models on task performance, inference cost, energy use and maintainability.

D:

For the decision, this would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, a larger model is not automatically the best option when the required task is narrow and resources are constrained. The more direct step is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability

Question 627

The organization chooses to compare smaller and larger models on task performance, inference cost, energy use and maintainability for this situation: A factory deploys visual defect detection at high speed on a constrained edge device. Which rationale best supports that choice?

  1. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  2. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  3. A larger model is not automatically the best option when the required task is narrow and resources are constrained.
  4. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Explanation

Correct Answer

C:

For the technical explanation, a larger model is not automatically the best option when the required task is narrow and resources are constrained. That is why the case calls for this action: compare smaller and larger models on task performance, inference cost, energy use and maintainability

Incorrect Answers

A:

For the technical explanation, that explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained

B:

For the technical explanation, that explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained

D:

That explains why a team might review whether an AI system is being repurposed for decisions, people or data beyond its approved scope, but it does not justify the decision required by the case. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Question 628

An AI governance reviewer examines this situation: A factory deploys visual defect detection at high speed on a constrained edge device. Which acceptance check is most decisive?

  1. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Observe simulated decisions with trained users and record comprehension, override and escalation results.
  3. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Explanation

Correct Answer

D:

For the verification, evaluate equivalent examples on accuracy, latency, operating expense and security requirements. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A larger model is not automatically the best option when the required task is narrow and resources are constrained

Incorrect Answers

A:

For the verification, document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements

B:

Observe simulated decisions with trained users and record comprehension, override and escalation results would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

C:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Question 629

A factory deploys visual defect detection at high speed on a constrained edge device. The team instead proposes: Select a massive general-purpose model solely because its parameter count is highest. Which failure should the reviewer expect most directly?

  1. Cost and latency can rise without meaningful gains for the actual workflow.
  2. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  3. Answers can become stale and lack traceable sources after the source policy changes.
  4. Model harms and wrong decisions can rise while infrastructure remains healthy.

Explanation

Correct Answer

A:

For the risk comparison, cost and latency can rise without meaningful gains for the actual workflow. The shortcut overlooks the central safeguard: compare smaller and larger models on task performance, inference cost, energy use and maintainability. A larger model is not automatically the best option when the required task is narrow and resources are constrained

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

C:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

D:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: Cost and latency can rise without meaningful gains for the actual workflow.

Question 630

A factory deploys visual defect detection at high speed on a constrained edge device. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  4. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.

Explanation

Correct Answer

B:

For the action and verification, compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements. Together, they implement and test the intended safeguard. A larger model is not automatically the best option when the required task is narrow and resources are constrained

Incorrect Answers

A:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

C:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

D:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. The supporting evidence is: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Question 631

A hospital introduces a triage assistant without training nurses on limitations and escalation. Which step would provide the most targeted control for the problem described?

  1. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.
  2. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  3. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  4. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Explanation

Correct Answer

C:

Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs. The practical acceptance evidence is: observe simulated decisions with trained users and record comprehension, override and escalation results.

Incorrect Answers

A:

The alternative—test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit—would not resolve the stated issue. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

B:

The alternative—compare smaller and larger models on task performance, inference cost, energy use and maintainability—would not resolve the stated issue. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

D:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs. The more direct step is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.

Question 632

A team reviewing the following situation asks for the most relevant governance principle: A hospital introduces a triage assistant without training nurses on limitations and escalation. Which principle best explains the required control?

  1. Natural-language generation and structured classification solve different problems and need different evaluations.
  2. A larger model is not automatically the best option when the required task is narrow and resources are constrained.
  3. Governance needs a response path for unsafe outcomes as well as conventional outages.
  4. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Explanation

Correct Answer

D:

Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs. That is why the case calls for this action: evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.

Incorrect Answers

A:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

B:

That explains why a team might compare smaller and larger models on task performance, inference cost, energy use and maintainability, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

C:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Question 633

An AI governance reviewer examines this situation: A hospital introduces a triage assistant without training nurses on limitations and escalation. Which acceptance check is most decisive?

  1. Observe simulated decisions with trained users and record comprehension, override and escalation results.
  2. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  3. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  4. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Explanation

Correct Answer

A:

Observe simulated decisions with trained users and record comprehension, override and escalation results. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Incorrect Answers

B:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results.

C:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results.

D:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results.

Question 634

The organization has this problem: A hospital introduces a triage assistant without training nurses on limitations and escalation. It takes a shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. What is the principal risk?

  1. Answers can become stale and lack traceable sources after the source policy changes.
  2. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  3. An agent may make unapproved purchases, record changes or external communications.
  4. Stakeholders may be misled and unable to take appropriate protective steps.

Explanation

Correct Answer

B:

Staff may over-rely on output or fail to recognize unsafe behavior in time. The shortcut overlooks the central safeguard: evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time.

C:

This consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time.

D:

This consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time.

Question 635

A hospital introduces a triage assistant without training nurses on limitations and escalation. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  2. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  3. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  4. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

C:

Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Observe simulated decisions with trained users and record comprehension, override and escalation results. Together, they implement and test the intended safeguard. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Incorrect Answers

A:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

B:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

D:

This pair supports another governance intervention: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

Question 636

Consider the following proposed AI use: A regional insurer replaces manual claim review screens before teaching adjusters how to challenge scores. Which response best addresses the governance gap?

  1. Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints.
  2. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  3. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.
  4. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.

Explanation

Correct Answer

D:

For the decision, evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs. The practical acceptance evidence is: observe simulated decisions with trained users and record comprehension, override and escalation results

Incorrect Answers

A:

The alternative—assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints—would not resolve the stated issue. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

B:

The alternative—specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy—would not resolve the stated issue. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

C:

The alternative—define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents—would not resolve the stated issue. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Question 637

A team reviewing the following situation asks for the most relevant governance principle: A regional insurer replaces manual claim review screens before teaching adjusters how to challenge scores. Which principle best explains the required control?

  1. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.
  2. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.
  3. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.
  4. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Explanation

Correct Answer

A:

For the technical explanation, even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs. That is why the case calls for this action: evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow

Incorrect Answers

B:

That explains why a team might specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

C:

That explains why a team might assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

D:

That explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Question 638

An AI governance reviewer examines this situation: A regional insurer replaces manual claim review screens before teaching adjusters how to challenge scores. Which acceptance check is most decisive?

  1. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  2. Observe simulated decisions with trained users and record comprehension, override and escalation results.
  3. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  4. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Explanation

Correct Answer

B:

For the verification, observe simulated decisions with trained users and record comprehension, override and escalation results. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs

Incorrect Answers

A:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results.

C:

For the verification, test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results

D:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches would test a different outcome. The directly relevant evidence is to observe simulated decisions with trained users and record comprehension, override and escalation results.

Question 639

A team considers this AI use: A regional insurer replaces manual claim review screens before teaching adjusters how to challenge scores. Rather than applying the relevant control, it will activate high-impact AI decisions without preparing the people expected to oversee them. What can go wrong?

  1. Connectivity and data-handling requirements may make a technically capable model unusable.
  2. Cost and latency can rise without meaningful gains for the actual workflow.
  3. Staff may over-rely on output or fail to recognize unsafe behavior in time.
  4. Stakeholders may be misled and unable to take appropriate protective steps.

Explanation

Correct Answer

C:

For the risk comparison, staff may over-rely on output or fail to recognize unsafe behavior in time. The shortcut overlooks the central safeguard: evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time.

B:

This consequence follows more directly from another shortcut: Select a massive general-purpose model solely because its parameter count is highest. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Staff may over-rely on output or fail to recognize unsafe behavior in time

Question 640

A regional insurer replaces manual claim review screens before teaching adjusters how to challenge scores. Which action and evidence pair most directly addresses the situation?

  1. Action: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. Evidence: Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  2. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  3. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.

Explanation

Correct Answer

D:

For the action and verification, evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Observe simulated decisions with trained users and record comprehension, override and escalation results. Together, they implement and test the intended safeguard. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs

Incorrect Answers

A:

This pair supports another governance intervention: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

B:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

C:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. The supporting evidence is: Observe simulated decisions with trained users and record comprehension, override and escalation results.

Question 641

An AI project team faces this situation: A municipal welfare office adopts an AI eligibility screener tested in retail customer service. What should the team do first to manage the central issue?

  1. Assess actual affected users, decisions, data and harms in the intended deployment environment.
  2. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  3. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  4. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Explanation

Correct Answer

A:

Assess actual affected users, decisions, data and harms in the intended deployment environment. A generic supplier assessment may not describe the deployment organization's population or operating conditions. The practical acceptance evidence is: compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Incorrect Answers

B:

This would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment.

C:

This would help address a separate governance task: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment.

D:

This would help address a separate governance task: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment.

Question 642

A municipal welfare office adopts an AI eligibility screener tested in retail customer service. Which explanation best connects the risk to the proposed governance action?

  1. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  2. A generic supplier assessment may not describe the deployment organization's population or operating conditions.
  3. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  4. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Explanation

Correct Answer

B:

A generic supplier assessment may not describe the deployment organization's population or operating conditions. That is why the case calls for this action: assess actual affected users, decisions, data and harms in the intended deployment environment.

Incorrect Answers

A:

That explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

C:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

D:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Question 643

A municipal welfare office adopts an AI eligibility screener tested in retail customer service. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Trace each included model and dataset license to the planned product's commercial and distribution use.
  2. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  3. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  4. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Explanation

Correct Answer

C:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Incorrect Answers

A:

Trace each included model and dataset license to the planned product's commercial and distribution use would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

B:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

D:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Question 644

For this case—A municipal welfare office adopts an AI eligibility screener tested in retail customer service—staff choose the shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. Which consequence best explains why this is inadequate?

  1. The system may fail on the very cases the deploying organization must handle.
  2. The deployment may violate license conditions or require unexpected changes.
  3. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.
  4. Site-specific harms and legal obligations may remain unrecognized.

Explanation

Correct Answer

D:

Site-specific harms and legal obligations may remain unrecognized. The shortcut overlooks the central safeguard: assess actual affected users, decisions, data and harms in the intended deployment environment. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized.

B:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized.

C:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized.

Question 645

The governance team needs both a decision and a way to verify it for this case: A municipal welfare office adopts an AI eligibility screener tested in retail customer service. Which pair is best?

  1. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  3. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  4. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Explanation

Correct Answer

A:

Assess actual affected users, decisions, data and harms in the intended deployment environment. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff. Together, they implement and test the intended safeguard. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Incorrect Answers

B:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

C:

This pair supports another governance intervention: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

D:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Question 646

An AI project team faces this situation: A university deploys a generic risk model for scholarship decisions without assessing student impacts. What should the team do first to manage the central issue?

  1. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  2. Assess actual affected users, decisions, data and harms in the intended deployment environment.
  3. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  4. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Explanation

Correct Answer

B:

For the decision, assess actual affected users, decisions, data and harms in the intended deployment environment. A generic supplier assessment may not describe the deployment organization's population or operating conditions. The practical acceptance evidence is: compare local impact assessment findings with vendor claims and obtain risk-owner signoff

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment

C:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment.

D:

This would help address a separate governance task: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For this case, however, a generic supplier assessment may not describe the deployment organization's population or operating conditions. The more direct step is: Assess actual affected users, decisions, data and harms in the intended deployment environment.

Question 647

A university deploys a generic risk model for scholarship decisions without assessing student impacts. Why is the recommended governance approach appropriate?

  1. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  2. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  3. A generic supplier assessment may not describe the deployment organization's population or operating conditions.
  4. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Explanation

Correct Answer

C:

For the technical explanation, a generic supplier assessment may not describe the deployment organization's population or operating conditions. That is why the case calls for this action: assess actual affected users, decisions, data and harms in the intended deployment environment

Incorrect Answers

A:

For the technical explanation, that explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions

B:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

D:

That explains why a team might compare smaller and larger models on task performance, inference cost, energy use and maintainability, but it does not justify the decision required by the case. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Question 648

A university deploys a generic risk model for scholarship decisions without assessing student impacts. Which recorded test or observation most directly demonstrates an effective response?

  1. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  2. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  3. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.
  4. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

D:

For the verification, compare local impact assessment findings with vendor claims and obtain risk-owner signoff. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A generic supplier assessment may not describe the deployment organization's population or operating conditions

Incorrect Answers

A:

Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

B:

For the verification, check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff

C:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Question 649

A university deploys a generic risk model for scholarship decisions without assessing student impacts. The team instead proposes: Reuse a supplier's generic impact assessment unchanged for a materially different local use. Which failure should the reviewer expect most directly?

  1. Site-specific harms and legal obligations may remain unrecognized.
  2. The deployment may violate license conditions or require unexpected changes.
  3. Investigations and regulatory responses may be delayed by missing evidence rights.
  4. Staff may over-rely on output or fail to recognize unsafe behavior in time.

Explanation

Correct Answer

A:

For the risk comparison, site-specific harms and legal obligations may remain unrecognized. The shortcut overlooks the central safeguard: assess actual affected users, decisions, data and harms in the intended deployment environment. A generic supplier assessment may not describe the deployment organization's population or operating conditions

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized

C:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized.

D:

This consequence follows more directly from another shortcut: Activate high-impact AI decisions without preparing the people expected to oversee them. The shortcut actually described would instead create this risk: Site-specific harms and legal obligations may remain unrecognized.

Question 650

Choose the intervention and validation that fit this business problem: A university deploys a generic risk model for scholarship decisions without assessing student impacts.

  1. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.
  2. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  3. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Explanation

Correct Answer

B:

For the action and verification, assess actual affected users, decisions, data and harms in the intended deployment environment. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff. Together, they implement and test the intended safeguard. A generic supplier assessment may not describe the deployment organization's population or operating conditions

Incorrect Answers

A:

This pair supports another governance intervention: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

C:

For the action and verification, this pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff

D:

This pair supports another governance intervention: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For the actual case, the required action is: Assess actual affected users, decisions, data and harms in the intended deployment environment. The supporting evidence is: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Question 651

Consider the following proposed AI use: A law firm evaluates an external document assistant for privileged client material. Which response best addresses the governance gap?

  1. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  2. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  3. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  4. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.

Explanation

Correct Answer

C:

Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Contractual restrictions determine whether supplied data may be retained or used for model improvement. The practical acceptance evidence is: check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Incorrect Answers

A:

This would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

B:

This would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

D:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Question 652

A law firm evaluates an external document assistant for privileged client material. Which explanation best connects the risk to the proposed governance action?

  1. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  2. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  3. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.
  4. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Explanation

Correct Answer

D:

Contractual restrictions determine whether supplied data may be retained or used for model improvement. That is why the case calls for this action: review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Incorrect Answers

A:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

B:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

C:

That explains why a team might maintain usable override and escalation mechanisms for consequential AI-assisted decisions, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Question 653

An AI governance reviewer examines this situation: A law firm evaluates an external document assistant for privileged client material. Which acceptance check is most decisive?

  1. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  2. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  3. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Explanation

Correct Answer

A:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Incorrect Answers

B:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls.

C:

Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls.

D:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Question 654

For this case—A law firm evaluates an external document assistant for privileged client material—staff choose the shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. Which consequence best explains why this is inadequate?

  1. The deployment may violate license conditions or require unexpected changes.
  2. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  3. Investigations and regulatory responses may be delayed by missing evidence rights.
  4. Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

Explanation

Correct Answer

B:

Sensitive inputs may be retained or reused beyond the organization's accepted boundaries. The shortcut overlooks the central safeguard: review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

C:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

D:

This consequence follows more directly from another shortcut: Switch off a legacy model without checking whether dependent workflows still call it or retain data. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

Question 655

A law firm evaluates an external document assistant for privileged client material. Which action and evidence pair most directly addresses the situation?

  1. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  2. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.
  3. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  4. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.

Explanation

Correct Answer

C:

Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Check signed terms and technical configuration for retention, reuse, deletion and processor controls. Together, they implement and test the intended safeguard. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

B:

This pair supports another governance intervention: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

D:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Question 656

Consider the following proposed AI use: An accounting firm enters payroll records into a hosted model without checking training-use terms. Which response best addresses the governance gap?

  1. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  2. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.
  3. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  4. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Explanation

Correct Answer

D:

For the decision, review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Contractual restrictions determine whether supplied data may be retained or used for model improvement. The practical acceptance evidence is: check signed terms and technical configuration for retention, reuse, deletion and processor controls

Incorrect Answers

A:

For the decision, this would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors

B:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

C:

This would help address a separate governance task: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For this case, however, contractual restrictions determine whether supplied data may be retained or used for model improvement. The more direct step is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Question 657

A team reviewing the following situation asks for the most relevant governance principle: An accounting firm enters payroll records into a hosted model without checking training-use terms. Which principle best explains the required control?

  1. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  2. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  3. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  4. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Explanation

Correct Answer

A:

For the technical explanation, contractual restrictions determine whether supplied data may be retained or used for model improvement. That is why the case calls for this action: review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors

Incorrect Answers

B:

That explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

C:

For the technical explanation, that explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement

D:

That explains why a team might implement and test a way to suspend unsafe AI functionality or switch to a safe fallback, but it does not justify the decision required by the case. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Question 658

An AI governance reviewer examines this situation: An accounting firm enters payroll records into a hosted model without checking training-use terms. Which acceptance check is most decisive?

  1. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  2. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  3. Trace each included model and dataset license to the planned product's commercial and distribution use.
  4. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Explanation

Correct Answer

B:

For the verification, check signed terms and technical configuration for retention, reuse, deletion and processor controls. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Contractual restrictions determine whether supplied data may be retained or used for model improvement

Incorrect Answers

A:

For the verification, compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls

C:

Trace each included model and dataset license to the planned product's commercial and distribution use would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls.

D:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Question 659

For this case—An accounting firm enters payroll records into a hosted model without checking training-use terms—staff choose the shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. Which consequence best explains why this is inadequate?

  1. The deployment may violate license conditions or require unexpected changes.
  2. The system may fail on the very cases the deploying organization must handle.
  3. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  4. The system may continue taking unsafe actions while teams debate a permanent correction.

Explanation

Correct Answer

C:

For the risk comparison, sensitive inputs may be retained or reused beyond the organization's accepted boundaries. The shortcut overlooks the central safeguard: review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Contractual restrictions determine whether supplied data may be retained or used for model improvement

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries

B:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

D:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

Question 660

The governance team needs both a decision and a way to verify it for this case: An accounting firm enters payroll records into a hosted model without checking training-use terms. Which pair is best?

  1. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.
  2. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  4. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Explanation

Correct Answer

D:

For the action and verification, review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Check signed terms and technical configuration for retention, reuse, deletion and processor controls. Together, they implement and test the intended safeguard. Contractual restrictions determine whether supplied data may be retained or used for model improvement

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls

B:

This pair supports another governance intervention: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

C:

This pair supports another governance intervention: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For the actual case, the required action is: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. The supporting evidence is: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.

Question 661

A financial services provider relies on a scoring vendor that will not disclose relevant incident records. Which step would provide the most targeted control for the problem described?

  1. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  2. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  3. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  4. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

Explanation

Correct Answer

A:

Define evidence access, incident notification and audit cooperation in third-party AI agreements. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The practical acceptance evidence is: test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Incorrect Answers

B:

This would help address a separate governance task: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

C:

This would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

D:

This would help address a separate governance task: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

Question 662

The organization chooses to define evidence access, incident notification and audit cooperation in third-party AI agreements for this situation: A financial services provider relies on a scoring vendor that will not disclose relevant incident records. Which rationale best supports that choice?

  1. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  2. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.
  3. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  4. Legal permission and local risk can depend on the place, population, processing and deployment context.

Explanation

Correct Answer

B:

A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. That is why the case calls for this action: define evidence access, incident notification and audit cooperation in third-party AI agreements.

Incorrect Answers

A:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

C:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

D:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Question 663

The organization has implemented a control for this case: A financial services provider relies on a scoring vendor that will not disclose relevant incident records. Which result would best validate it?

  1. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  3. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  4. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

C:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Incorrect Answers

A:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

B:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

D:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Question 664

A team considers this AI use: A financial services provider relies on a scoring vendor that will not disclose relevant incident records. Rather than applying the relevant control, it will assume the supplier will voluntarily share all incident logs despite no agreed process. What can go wrong?

  1. The deployment may violate license conditions or require unexpected changes.
  2. The organization can underbudget safety, data and operations obligations it now owns.
  3. A user can receive restricted information or trigger an unauthorized operation.
  4. Investigations and regulatory responses may be delayed by missing evidence rights.

Explanation

Correct Answer

D:

Investigations and regulatory responses may be delayed by missing evidence rights. The shortcut overlooks the central safeguard: define evidence access, incident notification and audit cooperation in third-party AI agreements. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

B:

This consequence follows more directly from another shortcut: Assume an in-house model removes vendor risk and therefore needs fewer governance controls. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

C:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

Question 665

A financial services provider relies on a scoring vendor that will not disclose relevant incident records. Which action and evidence pair most directly addresses the situation?

  1. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  2. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  3. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.
  4. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Explanation

Correct Answer

A:

Define evidence access, incident notification and audit cooperation in third-party AI agreements. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise. Together, they implement and test the intended safeguard. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Incorrect Answers

B:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

C:

This pair supports another governance intervention: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

D:

This pair supports another governance intervention: Monitor quality, safety, latency and feedback against meaningful production thresholds. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Question 666

Consider the following proposed AI use: A city agency's chatbot vendor provides no agreed notification channel for safety failures. Which response best addresses the governance gap?

  1. Assess actual affected users, decisions, data and harms in the intended deployment environment.
  2. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  3. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.
  4. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.

Explanation

Correct Answer

B:

For the decision, define evidence access, incident notification and audit cooperation in third-party AI agreements. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The practical acceptance evidence is: test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise

Incorrect Answers

A:

This would help address a separate governance task: Assess actual affected users, decisions, data and harms in the intended deployment environment. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

C:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

D:

This would help address a separate governance task: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For this case, however, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. The more direct step is: Define evidence access, incident notification and audit cooperation in third-party AI agreements.

Question 667

A city agency's chatbot vendor provides no agreed notification channel for safety failures. Which explanation best connects the risk to the proposed governance action?

  1. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  2. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  3. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.
  4. A model's own wording cannot substitute for external authorization checks on data and actions.

Explanation

Correct Answer

C:

For the technical explanation, a deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence. That is why the case calls for this action: define evidence access, incident notification and audit cooperation in third-party AI agreements

Incorrect Answers

A:

For the technical explanation, that explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence

B:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

D:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Question 668

An AI governance reviewer examines this situation: A city agency's chatbot vendor provides no agreed notification channel for safety failures. Which acceptance check is most decisive?

  1. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  3. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Explanation

Correct Answer

D:

For the verification, test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence

Incorrect Answers

A:

For the verification, compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise

B:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

C:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Question 669

The organization has this problem: A city agency's chatbot vendor provides no agreed notification channel for safety failures. It takes a shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. What is the principal risk?

  1. Investigations and regulatory responses may be delayed by missing evidence rights.
  2. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  3. Site-specific harms and legal obligations may remain unrecognized.
  4. The system may perform restricted activities in places where different rules apply.

Explanation

Correct Answer

A:

For the risk comparison, investigations and regulatory responses may be delayed by missing evidence rights. The shortcut overlooks the central safeguard: define evidence access, incident notification and audit cooperation in third-party AI agreements. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

C:

This consequence follows more directly from another shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

D:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: Investigations and regulatory responses may be delayed by missing evidence rights.

Question 670

The governance team needs both a decision and a way to verify it for this case: A city agency's chatbot vendor provides no agreed notification channel for safety failures. Which pair is best?

  1. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  2. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  4. Action: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Evidence: Test users with different roles against protected documents and disallowed tool operations.

Explanation

Correct Answer

B:

For the action and verification, define evidence access, incident notification and audit cooperation in third-party AI agreements. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise. Together, they implement and test the intended safeguard. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

C:

This pair supports another governance intervention: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

D:

This pair supports another governance intervention: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For the actual case, the required action is: Define evidence access, incident notification and audit cooperation in third-party AI agreements. The supporting evidence is: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Question 671

Consider the following proposed AI use: A startup embeds an openly published model in a paid on-device product without checking its license. Which response best addresses the governance gap?

  1. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  2. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  3. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  4. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.

Explanation

Correct Answer

C:

Verify model, dataset and output licensing conditions before integrating or redistributing AI components. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The practical acceptance evidence is: trace each included model and dataset license to the planned product's commercial and distribution use.

Incorrect Answers

A:

This would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

B:

This would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

D:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Question 672

A startup embeds an openly published model in a paid on-device product without checking its license. Why is the recommended governance approach appropriate?

  1. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  2. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  3. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.
  4. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Explanation

Correct Answer

D:

An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. That is why the case calls for this action: verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Incorrect Answers

A:

That explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

B:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

C:

That explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Question 673

A startup embeds an openly published model in a paid on-device product without checking its license. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Trace each included model and dataset license to the planned product's commercial and distribution use.
  2. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  4. Audit overridden cases for decision context, reviewer authority and follow-through.

Explanation

Correct Answer

A:

Trace each included model and dataset license to the planned product's commercial and distribution use. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Incorrect Answers

B:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use.

C:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use.

D:

Audit overridden cases for decision context, reviewer authority and follow-through would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use.

Question 674

A team considers this AI use: A startup embeds an openly published model in a paid on-device product without checking its license. Rather than applying the relevant control, it will describe every openly downloadable model as unrestricted public-domain software. What can go wrong?

  1. The organization can underbudget safety, data and operations obligations it now owns.
  2. The deployment may violate license conditions or require unexpected changes.
  3. Investigations and regulatory responses may be delayed by missing evidence rights.
  4. Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

Explanation

Correct Answer

B:

The deployment may violate license conditions or require unexpected changes. The shortcut overlooks the central safeguard: verify model, dataset and output licensing conditions before integrating or redistributing AI components. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume an in-house model removes vendor risk and therefore needs fewer governance controls. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

C:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

D:

This consequence follows more directly from another shortcut: Switch off a legacy model without checking whether dependent workflows still call it or retain data. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

Question 675

A startup embeds an openly published model in a paid on-device product without checking its license. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.
  4. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Explanation

Correct Answer

C:

Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Trace each included model and dataset license to the planned product's commercial and distribution use. Together, they implement and test the intended safeguard. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Incorrect Answers

A:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use.

B:

This pair supports another governance intervention: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use.

D:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use.

Question 676

A research consortium offers a redistributed fine-tuned model without reviewing base model restrictions. Which governance action most directly addresses the need in this situation?

  1. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  2. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  3. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.
  4. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Explanation

Correct Answer

D:

For the decision, verify model, dataset and output licensing conditions before integrating or redistributing AI components. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The practical acceptance evidence is: trace each included model and dataset license to the planned product's commercial and distribution use

Incorrect Answers

A:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

B:

For the decision, this would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components

C:

This would help address a separate governance task: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For this case, however, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. The more direct step is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Question 677

A research consortium offers a redistributed fine-tuned model without reviewing base model restrictions. Why is the recommended governance approach appropriate?

  1. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  2. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.
  3. A generic supplier assessment may not describe the deployment organization's population or operating conditions.
  4. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Explanation

Correct Answer

A:

For the technical explanation, an open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions. That is why the case calls for this action: verify model, dataset and output licensing conditions before integrating or redistributing AI components

Incorrect Answers

B:

That explains why a team might define evidence access, incident notification and audit cooperation in third-party AI agreements, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

C:

That explains why a team might assess actual affected users, decisions, data and harms in the intended deployment environment, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

D:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Question 678

A research consortium offers a redistributed fine-tuned model without reviewing base model restrictions. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  2. Trace each included model and dataset license to the planned product's commercial and distribution use.
  3. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  4. Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

B:

For the verification, trace each included model and dataset license to the planned product's commercial and distribution use. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions

Incorrect Answers

A:

For the verification, test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use

C:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use.

D:

Run representative task trials measuring utility, error types, explainability and required operating controls would test a different outcome. The directly relevant evidence is to trace each included model and dataset license to the planned product's commercial and distribution use.

Question 679

A research consortium offers a redistributed fine-tuned model without reviewing base model restrictions. The team instead proposes: Describe every openly downloadable model as unrestricted public-domain software. Which failure should the reviewer expect most directly?

  1. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  2. The system may fail on the very cases the deploying organization must handle.
  3. The deployment may violate license conditions or require unexpected changes.
  4. A user can receive restricted information or trigger an unauthorized operation.

Explanation

Correct Answer

C:

For the risk comparison, the deployment may violate license conditions or require unexpected changes. The shortcut overlooks the central safeguard: verify model, dataset and output licensing conditions before integrating or redistributing AI components. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

B:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

D:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: The deployment may violate license conditions or require unexpected changes.

Question 680

A research consortium offers a redistributed fine-tuned model without reviewing base model restrictions. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.
  3. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.

Explanation

Correct Answer

D:

For the action and verification, verify model, dataset and output licensing conditions before integrating or redistributing AI components. Trace each included model and dataset license to the planned product's commercial and distribution use. Together, they implement and test the intended safeguard. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use

B:

This pair supports another governance intervention: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use.

C:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. The supporting evidence is: Trace each included model and dataset license to the planned product's commercial and distribution use.

Question 681

A large retailer proposes training its own recommendation model on mixed customer data to avoid external APIs. Which step would provide the most targeted control for the problem described?

  1. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  2. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.
  3. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  4. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.

Explanation

Correct Answer

A:

Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. The practical acceptance evidence is: review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Incorrect Answers

B:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. The more direct step is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.

C:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. The more direct step is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.

D:

The alternative—test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit—would not resolve the stated issue. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Question 682

A large retailer proposes training its own recommendation model on mixed customer data to avoid external APIs. Which explanation best connects the risk to the proposed governance action?

  1. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  2. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  3. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  4. Tool-using systems can create side effects, so a plausible language response is not the same as authorization to act.

Explanation

Correct Answer

B:

Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. That is why the case calls for this action: evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.

Incorrect Answers

A:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

C:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

D:

That explains why a team might specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Question 683

A large retailer proposes training its own recommendation model on mixed customer data to avoid external APIs. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  2. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  3. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Explanation

Correct Answer

C:

Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Incorrect Answers

A:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

B:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

D:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Question 684

A team considers this AI use: A large retailer proposes training its own recommendation model on mixed customer data to avoid external APIs. Rather than applying the relevant control, it will assume an in-house model removes vendor risk and therefore needs fewer governance controls. What can go wrong?

  1. The deployment may violate license conditions or require unexpected changes.
  2. Investigations and regulatory responses may be delayed by missing evidence rights.
  3. Cost and latency can rise without meaningful gains for the actual workflow.
  4. The organization can underbudget safety, data and operations obligations it now owns.

Explanation

Correct Answer

D:

The organization can underbudget safety, data and operations obligations it now owns. The shortcut overlooks the central safeguard: evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

B:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

C:

This consequence follows more directly from another shortcut: Select a massive general-purpose model solely because its parameter count is highest. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

Question 685

Choose the intervention and validation that fit this business problem: A large retailer proposes training its own recommendation model on mixed customer data to avoid external APIs.

  1. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  2. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  3. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  4. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Explanation

Correct Answer

A:

Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities. Together, they implement and test the intended safeguard. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Incorrect Answers

B:

This pair supports another governance intervention: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

C:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

D:

This pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Question 686

A telecommunications company considers building a language model without a dedicated post-release safety team. Which step would provide the most targeted control for the problem described?

  1. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.
  2. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  3. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.
  4. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Explanation

Correct Answer

B:

For the decision, evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. The practical acceptance evidence is: review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities

Incorrect Answers

A:

The alternative—verify model, dataset and output licensing conditions before integrating or redistributing AI components—would not resolve the stated issue. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

C:

For the decision, this would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. The more direct step is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model

D:

The alternative—implement and test a way to suspend unsafe AI functionality or switch to a safe fallback—would not resolve the stated issue. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Question 687

A team reviewing the following situation asks for the most relevant governance principle: A telecommunications company considers building a language model without a dedicated post-release safety team. Which principle best explains the required control?

  1. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  2. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  3. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  4. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Explanation

Correct Answer

C:

For the technical explanation, owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations. That is why the case calls for this action: evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model

Incorrect Answers

A:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

B:

For the technical explanation, that explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations

D:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.

Question 688

The organization has implemented a control for this case: A telecommunications company considers building a language model without a dedicated post-release safety team. Which result would best validate it?

  1. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  2. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  3. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  4. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Explanation

Correct Answer

D:

For the verification, review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations

Incorrect Answers

A:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

B:

Compare supplier claims with local performance, bias and robustness tests using representative samples would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

C:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check would test a different outcome. The directly relevant evidence is to review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Question 689

A team considers this AI use: A telecommunications company considers building a language model without a dedicated post-release safety team. Rather than applying the relevant control, it will assume an in-house model removes vendor risk and therefore needs fewer governance controls. What can go wrong?

  1. The organization can underbudget safety, data and operations obligations it now owns.
  2. The system may fail on the very cases the deploying organization must handle.
  3. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  4. Model harms and wrong decisions can rise while infrastructure remains healthy.

Explanation

Correct Answer

A:

For the risk comparison, the organization can underbudget safety, data and operations obligations it now owns. The shortcut overlooks the central safeguard: evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

C:

This consequence follows more directly from another shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

D:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: The organization can underbudget safety, data and operations obligations it now owns.

Question 690

A telecommunications company considers building a language model without a dedicated post-release safety team. Which action and evidence pair most directly addresses the situation?

  1. Action: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. Evidence: Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  2. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  3. Action: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. Evidence: Trace each included model and dataset license to the planned product's commercial and distribution use.
  4. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Explanation

Correct Answer

B:

For the action and verification, evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities. Together, they implement and test the intended safeguard. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities

C:

This pair supports another governance intervention: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

D:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. The supporting evidence is: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Question 691

Consider the following proposed AI use: A regional hospital buys a voice tool tested only in one language but serves multilingual callers. Which response best addresses the governance gap?

  1. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  2. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.
  3. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.
  4. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

Explanation

Correct Answer

C:

Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Benchmark results obtained under another population, language or workflow may not transfer to local use. The practical acceptance evidence is: compare supplier claims with local performance, bias and robustness tests using representative samples.

Incorrect Answers

A:

This would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

B:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

D:

This would help address a separate governance task: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Question 692

The organization chooses to run deployment-representative tests instead of accepting supplier performance claims as sufficient proof for this situation: A regional hospital buys a voice tool tested only in one language but serves multilingual callers. Which rationale best supports that choice?

  1. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  2. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  3. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  4. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Explanation

Correct Answer

D:

Benchmark results obtained under another population, language or workflow may not transfer to local use. That is why the case calls for this action: run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Incorrect Answers

A:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use.

B:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use.

C:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Question 693

The organization has implemented a control for this case: A regional hospital buys a voice tool tested only in one language but serves multilingual callers. Which result would best validate it?

  1. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  2. Trace each included model and dataset license to the planned product's commercial and distribution use.
  3. Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Explanation

Correct Answer

A:

Compare supplier claims with local performance, bias and robustness tests using representative samples. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Incorrect Answers

B:

Trace each included model and dataset license to the planned product's commercial and distribution use would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

C:

Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

D:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

Question 694

For this case—A regional hospital buys a voice tool tested only in one language but serves multilingual callers—staff choose the shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. Which consequence best explains why this is inadequate?

  1. The deployment may violate license conditions or require unexpected changes.
  2. The system may fail on the very cases the deploying organization must handle.
  3. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.
  4. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Explanation

Correct Answer

B:

The system may fail on the very cases the deploying organization must handle. The shortcut overlooks the central safeguard: run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

C:

This consequence follows more directly from another shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

D:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

Question 695

A regional hospital buys a voice tool tested only in one language but serves multilingual callers. Which action and evidence pair most directly addresses the situation?

  1. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.
  2. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.
  4. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.

Explanation

Correct Answer

C:

Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Compare supplier claims with local performance, bias and robustness tests using representative samples. Together, they implement and test the intended safeguard. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Incorrect Answers

A:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples.

B:

This pair supports another governance intervention: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples.

D:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Question 696

Consider the following proposed AI use: A shipping insurer selects a damage classifier validated only on studio-quality product images. Which response best addresses the governance gap?

  1. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  2. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.
  3. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  4. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Explanation

Correct Answer

D:

For the decision, run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Benchmark results obtained under another population, language or workflow may not transfer to local use. The practical acceptance evidence is: compare supplier claims with local performance, bias and robustness tests using representative samples

Incorrect Answers

A:

This would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

B:

For the decision, this would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof

C:

This would help address a separate governance task: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For this case, however, benchmark results obtained under another population, language or workflow may not transfer to local use. The more direct step is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Question 697

A shipping insurer selects a damage classifier validated only on studio-quality product images. Why is the recommended governance approach appropriate?

  1. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  2. A generic supplier assessment may not describe the deployment organization's population or operating conditions.
  3. Contractual restrictions determine whether supplied data may be retained or used for model improvement.
  4. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Explanation

Correct Answer

A:

For the technical explanation, benchmark results obtained under another population, language or workflow may not transfer to local use. That is why the case calls for this action: run deployment-representative tests instead of accepting supplier performance claims as sufficient proof

Incorrect Answers

B:

That explains why a team might assess actual affected users, decisions, data and harms in the intended deployment environment, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use.

C:

For the technical explanation, that explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use

D:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Question 698

An AI governance reviewer examines this situation: A shipping insurer selects a damage classifier validated only on studio-quality product images. Which acceptance check is most decisive?

  1. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  2. Compare supplier claims with local performance, bias and robustness tests using representative samples.
  3. Check signed terms and technical configuration for retention, reuse, deletion and processor controls.
  4. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

B:

For the verification, compare supplier claims with local performance, bias and robustness tests using representative samples. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Benchmark results obtained under another population, language or workflow may not transfer to local use

Incorrect Answers

A:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

C:

Check signed terms and technical configuration for retention, reuse, deletion and processor controls would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

D:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to compare supplier claims with local performance, bias and robustness tests using representative samples.

Question 699

A team considers this AI use: A shipping insurer selects a damage classifier validated only on studio-quality product images. Rather than applying the relevant control, it will approve a model because the supplier reports excellent global benchmark accuracy without local testing. What can go wrong?

  1. Site-specific harms and legal obligations may remain unrecognized.
  2. Investigations and regulatory responses may be delayed by missing evidence rights.
  3. The system may fail on the very cases the deploying organization must handle.
  4. Staff can act on fabricated answers or expose sensitive information through normal workflows.

Explanation

Correct Answer

C:

For the risk comparison, the system may fail on the very cases the deploying organization must handle. The shortcut overlooks the central safeguard: run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Benchmark results obtained under another population, language or workflow may not transfer to local use

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

B:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

D:

This consequence follows more directly from another shortcut: Treat access to an AI tool as proof that employees know how to use it safely. The shortcut actually described would instead create this risk: The system may fail on the very cases the deploying organization must handle.

Question 700

Choose the intervention and validation that fit this business problem: A shipping insurer selects a damage classifier validated only on studio-quality product images.

  1. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  2. Action: Define evidence access, incident notification and audit cooperation in third-party AI agreements. Evidence: Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  3. Action: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. Evidence: Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

D:

For the action and verification, run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Compare supplier claims with local performance, bias and robustness tests using representative samples. Together, they implement and test the intended safeguard. Benchmark results obtained under another population, language or workflow may not transfer to local use

Incorrect Answers

A:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples.

B:

For the action and verification, this pair supports another governance intervention: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples

C:

This pair supports another governance intervention: Assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints. For the actual case, the required action is: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. The supporting evidence is: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Question 701

Consider the following proposed AI use: A legal assistant can search personnel files outside the requesting attorney's role. Which response best addresses the governance gap?

  1. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  2. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  3. Use region- and role-aware controls when the permitted AI use differs across operating locations.
  4. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Explanation

Correct Answer

A:

Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. A model's own wording cannot substitute for external authorization checks on data and actions. The practical acceptance evidence is: test users with different roles against protected documents and disallowed tool operations.

Incorrect Answers

B:

This would help address a separate governance task: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

C:

This would help address a separate governance task: Use region- and role-aware controls when the permitted AI use differs across operating locations. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

D:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

Question 702

A team reviewing the following situation asks for the most relevant governance principle: A legal assistant can search personnel files outside the requesting attorney's role. Which principle best explains the required control?

  1. Trust and accountability depend on timely, factual communication of material changes and significant issues.
  2. A model's own wording cannot substitute for external authorization checks on data and actions.
  3. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  4. Natural-language generation and structured classification solve different problems and need different evaluations.

Explanation

Correct Answer

B:

A model's own wording cannot substitute for external authorization checks on data and actions. That is why the case calls for this action: apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

Incorrect Answers

A:

That explains why a team might develop an accurate incident and capability communication plan for users, customers and external stakeholders, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

C:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

D:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

Question 703

A legal assistant can search personnel files outside the requesting attorney's role. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  2. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  3. Test users with different roles against protected documents and disallowed tool operations.
  4. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.

Explanation

Correct Answer

C:

Test users with different roles against protected documents and disallowed tool operations. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A model's own wording cannot substitute for external authorization checks on data and actions.

Incorrect Answers

A:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

B:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

D:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

Question 704

The organization has this problem: A legal assistant can search personnel files outside the requesting attorney's role. It takes a shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. What is the principal risk?

  1. The system may continue taking unsafe actions while teams debate a permanent correction.
  2. Model harms and wrong decisions can rise while infrastructure remains healthy.
  3. The deployment may violate license conditions or require unexpected changes.
  4. A user can receive restricted information or trigger an unauthorized operation.

Explanation

Correct Answer

D:

A user can receive restricted information or trigger an unauthorized operation. The shortcut overlooks the central safeguard: apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. A model's own wording cannot substitute for external authorization checks on data and actions.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

B:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

C:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

Question 705

Choose the intervention and validation that fit this business problem: A legal assistant can search personnel files outside the requesting attorney's role.

  1. Action: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Evidence: Test users with different roles against protected documents and disallowed tool operations.
  2. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  3. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.

Explanation

Correct Answer

A:

Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Test users with different roles against protected documents and disallowed tool operations. Together, they implement and test the intended safeguard. A model's own wording cannot substitute for external authorization checks on data and actions.

Incorrect Answers

B:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations.

C:

This pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations.

D:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations.

Question 706

Consider the following proposed AI use: An employee agent can open another team's payroll documents despite ordinary application restrictions. Which response best addresses the governance gap?

  1. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.
  2. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  3. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  4. Verify model, dataset and output licensing conditions before integrating or redistributing AI components.

Explanation

Correct Answer

B:

For the decision, apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. A model's own wording cannot substitute for external authorization checks on data and actions. The practical acceptance evidence is: test users with different roles against protected documents and disallowed tool operations

Incorrect Answers

A:

This would help address a separate governance task: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

C:

This would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

D:

This would help address a separate governance task: Verify model, dataset and output licensing conditions before integrating or redistributing AI components. For this case, however, a model's own wording cannot substitute for external authorization checks on data and actions. The more direct step is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.

Question 707

An employee agent can open another team's payroll documents despite ordinary application restrictions. Why is the recommended governance approach appropriate?

  1. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.
  2. A safe assessment for one use case does not establish suitability for a materially different downstream decision.
  3. A model's own wording cannot substitute for external authorization checks on data and actions.
  4. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Explanation

Correct Answer

C:

For the technical explanation, a model's own wording cannot substitute for external authorization checks on data and actions. That is why the case calls for this action: apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation

Incorrect Answers

A:

That explains why a team might implement and test a way to suspend unsafe AI functionality or switch to a safe fallback, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

B:

That explains why a team might review whether an AI system is being repurposed for decisions, people or data beyond its approved scope, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

D:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. A model's own wording cannot substitute for external authorization checks on data and actions.

Question 708

The organization has implemented a control for this case: An employee agent can open another team's payroll documents despite ordinary application restrictions. Which result would best validate it?

  1. Audit overridden cases for decision context, reviewer authority and follow-through.
  2. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  3. Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Test users with different roles against protected documents and disallowed tool operations.

Explanation

Correct Answer

D:

For the verification, test users with different roles against protected documents and disallowed tool operations. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A model's own wording cannot substitute for external authorization checks on data and actions

Incorrect Answers

A:

Audit overridden cases for decision context, reviewer authority and follow-through would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

B:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

C:

Run representative task trials measuring utility, error types, explainability and required operating controls would test a different outcome. The directly relevant evidence is to test users with different roles against protected documents and disallowed tool operations.

Question 709

For this case—An employee agent can open another team's payroll documents despite ordinary application restrictions—staff choose the shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. Which consequence best explains why this is inadequate?

  1. A user can receive restricted information or trigger an unauthorized operation.
  2. Unsafe behavior may continue without a responsible containment decision.
  3. Unassessed harms may arise from new decision contexts and affected groups.
  4. Investigations and regulatory responses may be delayed by missing evidence rights.

Explanation

Correct Answer

A:

For the risk comparison, a user can receive restricted information or trigger an unauthorized operation. The shortcut overlooks the central safeguard: apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. A model's own wording cannot substitute for external authorization checks on data and actions

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

C:

This consequence follows more directly from another shortcut: Treat approval for inventory forecasting as approval to score individual employees' productivity. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

D:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: A user can receive restricted information or trigger an unauthorized operation.

Question 710

Choose the intervention and validation that fit this business problem: An employee agent can open another team's payroll documents despite ordinary application restrictions.

  1. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  2. Action: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Evidence: Test users with different roles against protected documents and disallowed tool operations.
  3. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

B:

For the action and verification, apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Test users with different roles against protected documents and disallowed tool operations. Together, they implement and test the intended safeguard. A model's own wording cannot substitute for external authorization checks on data and actions

Incorrect Answers

A:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations.

C:

For the action and verification, this pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations

D:

This pair supports another governance intervention: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For the actual case, the required action is: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. The supporting evidence is: Test users with different roles against protected documents and disallowed tool operations.

Question 711

Consider the following proposed AI use: A public helpline AI stays online but increasingly gives outdated benefits information. Which response best addresses the governance gap?

  1. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Monitor quality, safety, latency and feedback against meaningful production thresholds.
  4. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.

Explanation

Correct Answer

C:

Monitor quality, safety, latency and feedback against meaningful production thresholds. Availability alone does not reveal harmful answers, performance drift or rising decision error rates. The practical acceptance evidence is: inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Incorrect Answers

A:

This would help address a separate governance task: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds.

B:

This would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds.

D:

This would help address a separate governance task: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds.

Question 712

A public helpline AI stays online but increasingly gives outdated benefits information. Which explanation best connects the risk to the proposed governance action?

  1. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  2. A model's own wording cannot substitute for external authorization checks on data and actions.
  3. Owning the full model stack provides control but also requires evidence for data rights, training, evaluation and ongoing operations.
  4. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Explanation

Correct Answer

D:

Availability alone does not reveal harmful answers, performance drift or rising decision error rates. That is why the case calls for this action: monitor quality, safety, latency and feedback against meaningful production thresholds.

Incorrect Answers

A:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

B:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

C:

That explains why a team might evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Question 713

The organization has implemented a control for this case: A public helpline AI stays online but increasingly gives outdated benefits information. Which result would best validate it?

  1. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  2. Run a response exercise and verify incident logging, containment actions and post-incident review.
  3. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  4. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Explanation

Correct Answer

A:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Incorrect Answers

B:

Run a response exercise and verify incident logging, containment actions and post-incident review would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

C:

Inspect retirement records for disabled access, archive obligations, migration checks and owner approval would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

D:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Question 714

The organization has this problem: A public helpline AI stays online but increasingly gives outdated benefits information. It takes a shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. What is the principal risk?

  1. Staff can act on fabricated answers or expose sensitive information through normal workflows.
  2. Model harms and wrong decisions can rise while infrastructure remains healthy.
  3. The system may perform restricted activities in places where different rules apply.
  4. Site-specific harms and legal obligations may remain unrecognized.

Explanation

Correct Answer

B:

Model harms and wrong decisions can rise while infrastructure remains healthy. The shortcut overlooks the central safeguard: monitor quality, safety, latency and feedback against meaningful production thresholds. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat access to an AI tool as proof that employees know how to use it safely. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy.

C:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy.

D:

This consequence follows more directly from another shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy.

Question 715

The governance team needs both a decision and a way to verify it for this case: A public helpline AI stays online but increasingly gives outdated benefits information. Which pair is best?

  1. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  3. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  4. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

C:

Monitor quality, safety, latency and feedback against meaningful production thresholds. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches. Together, they implement and test the intended safeguard. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Incorrect Answers

A:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

B:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

D:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Question 716

Consider the following proposed AI use: A manufacturing defect detector has perfect uptime but its missed-defect rate increases after a camera change. Which response best addresses the governance gap?

  1. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Define evidence access, incident notification and audit cooperation in third-party AI agreements.
  4. Monitor quality, safety, latency and feedback against meaningful production thresholds.

Explanation

Correct Answer

D:

For the decision, monitor quality, safety, latency and feedback against meaningful production thresholds. Availability alone does not reveal harmful answers, performance drift or rising decision error rates. The practical acceptance evidence is: inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches

Incorrect Answers

A:

This would help address a separate governance task: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds.

B:

For the decision, this would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds

C:

This would help address a separate governance task: Define evidence access, incident notification and audit cooperation in third-party AI agreements. For this case, however, availability alone does not reveal harmful answers, performance drift or rising decision error rates. The more direct step is: Monitor quality, safety, latency and feedback against meaningful production thresholds.

Question 717

The organization chooses to monitor quality, safety, latency and feedback against meaningful production thresholds for this situation: A manufacturing defect detector has perfect uptime but its missed-defect rate increases after a camera change. Which rationale best supports that choice?

  1. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  2. Governance needs a response path for unsafe outcomes as well as conventional outages.
  3. Legal permission and local risk can depend on the place, population, processing and deployment context.
  4. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.

Explanation

Correct Answer

A:

For the technical explanation, availability alone does not reveal harmful answers, performance drift or rising decision error rates. That is why the case calls for this action: monitor quality, safety, latency and feedback against meaningful production thresholds

Incorrect Answers

B:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

C:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

D:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.

Question 718

A manufacturing defect detector has perfect uptime but its missed-defect rate increases after a camera change. Which recorded test or observation most directly demonstrates an effective response?

  1. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  2. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  3. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  4. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Explanation

Correct Answer

B:

For the verification, inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Availability alone does not reveal harmful answers, performance drift or rising decision error rates

Incorrect Answers

A:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

C:

For the verification, inspect retirement records for disabled access, archive obligations, migration checks and owner approval would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches

D:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Question 719

A team considers this AI use: A manufacturing defect detector has perfect uptime but its missed-defect rate increases after a camera change. Rather than applying the relevant control, it will watch only server uptime as evidence that an AI-assisted service is performing acceptably. What can go wrong?

  1. Stakeholders may be misled and unable to take appropriate protective steps.
  2. A user can receive restricted information or trigger an unauthorized operation.
  3. Model harms and wrong decisions can rise while infrastructure remains healthy.
  4. Site-specific harms and legal obligations may remain unrecognized.

Explanation

Correct Answer

C:

For the risk comparison, model harms and wrong decisions can rise while infrastructure remains healthy. The shortcut overlooks the central safeguard: monitor quality, safety, latency and feedback against meaningful production thresholds. Availability alone does not reveal harmful answers, performance drift or rising decision error rates

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy.

B:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy.

D:

For the risk comparison, this consequence follows more directly from another shortcut: Reuse a supplier's generic impact assessment unchanged for a materially different local use. The shortcut actually described would instead create this risk: Model harms and wrong decisions can rise while infrastructure remains healthy

Question 720

The governance team needs both a decision and a way to verify it for this case: A manufacturing defect detector has perfect uptime but its missed-defect rate increases after a camera change. Which pair is best?

  1. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  2. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  3. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Explanation

Correct Answer

D:

For the action and verification, monitor quality, safety, latency and feedback against meaningful production thresholds. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches. Together, they implement and test the intended safeguard. Availability alone does not reveal harmful answers, performance drift or rising decision error rates

Incorrect Answers

A:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

B:

For the action and verification, this pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches

C:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Monitor quality, safety, latency and feedback against meaningful production thresholds. The supporting evidence is: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.

Question 721

A lending reviewer discovers an incorrect denial but cannot stop the automatic notice. Which governance action most directly addresses the need in this situation?

  1. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Monitor quality, safety, latency and feedback against meaningful production thresholds.
  4. Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof.

Explanation

Correct Answer

A:

Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The practical acceptance evidence is: audit overridden cases for decision context, reviewer authority and follow-through.

Incorrect Answers

B:

This would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

C:

This would help address a separate governance task: Monitor quality, safety, latency and feedback against meaningful production thresholds. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

D:

This would help address a separate governance task: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

Question 722

A lending reviewer discovers an incorrect denial but cannot stop the automatic notice. Why is the recommended governance approach appropriate?

  1. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  2. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.
  3. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  4. Retrieval can provide current sourced context without changing weights, while adaptation modifies model behavior and introduces separate upkeep needs.

Explanation

Correct Answer

B:

Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. That is why the case calls for this action: maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

Incorrect Answers

A:

That explains why a team might plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

C:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

D:

That explains why a team might test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Question 723

The organization has implemented a control for this case: A lending reviewer discovers an incorrect denial but cannot stop the automatic notice. Which result would best validate it?

  1. Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Test users with different roles against protected documents and disallowed tool operations.
  3. Audit overridden cases for decision context, reviewer authority and follow-through.
  4. Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

C:

Audit overridden cases for decision context, reviewer authority and follow-through. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Incorrect Answers

A:

The alternative—run a response exercise and verify incident logging, containment actions and post-incident review—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

B:

The alternative—test users with different roles against protected documents and disallowed tool operations—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

D:

The alternative—run representative task trials measuring utility, error types, explainability and required operating controls—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Question 724

A lending reviewer discovers an incorrect denial but cannot stop the automatic notice. The team instead proposes: Make the review button visible but technically prevent staff from changing an AI recommendation. Which failure should the reviewer expect most directly?

  1. The system may perform restricted activities in places where different rules apply.
  2. Staff can act on fabricated answers or expose sensitive information through normal workflows.
  3. Investigations and regulatory responses may be delayed by missing evidence rights.
  4. People may be unable to correct known errors in time to avoid harm.

Explanation

Correct Answer

D:

People may be unable to correct known errors in time to avoid harm. The shortcut overlooks the central safeguard: maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

B:

This consequence follows more directly from another shortcut: Treat access to an AI tool as proof that employees know how to use it safely. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

C:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

Question 725

A lending reviewer discovers an incorrect denial but cannot stop the automatic notice. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  2. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  3. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  4. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

A:

Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Audit overridden cases for decision context, reviewer authority and follow-through. Together, they implement and test the intended safeguard. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Incorrect Answers

B:

This pair supports another governance intervention: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

C:

This pair supports another governance intervention: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

D:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

Question 726

An AI project team faces this situation: A hospital triage nurse sees an unsafe priority score but has no clear escalation control. What should the team do first to manage the central issue?

  1. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  2. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  3. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.
  4. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.

Explanation

Correct Answer

B:

For the decision, maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The practical acceptance evidence is: audit overridden cases for decision context, reviewer authority and follow-through

Incorrect Answers

A:

This would help address a separate governance task: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

C:

This would help address a separate governance task: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

D:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. The more direct step is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.

Question 727

A hospital triage nurse sees an unsafe priority score but has no clear escalation control. Which explanation best connects the risk to the proposed governance action?

  1. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  2. Trust and accountability depend on timely, factual communication of material changes and significant issues.
  3. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.
  4. A larger model is not automatically the best option when the required task is narrow and resources are constrained.

Explanation

Correct Answer

C:

For the technical explanation, human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate. That is why the case calls for this action: maintain usable override and escalation mechanisms for consequential AI-assisted decisions

Incorrect Answers

A:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

B:

That explains why a team might develop an accurate incident and capability communication plan for users, customers and external stakeholders, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

D:

That explains why a team might compare smaller and larger models on task performance, inference cost, energy use and maintainability, but it does not justify the decision required by the case. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Question 728

The organization has implemented a control for this case: A hospital triage nurse sees an unsafe priority score but has no clear escalation control. Which result would best validate it?

  1. Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  3. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Audit overridden cases for decision context, reviewer authority and follow-through.

Explanation

Correct Answer

D:

For the verification, audit overridden cases for decision context, reviewer authority and follow-through. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate

Incorrect Answers

A:

For the verification, the alternative—run a response exercise and verify incident logging, containment actions and post-incident review—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate

B:

The alternative—test restricted-region access, user notices and fail-closed behavior against approved policy decisions—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

C:

The alternative—document a fit assessment with real latency, connectivity, privacy and maintenance tests—would not resolve the stated issue. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.

Question 729

A team considers this AI use: A hospital triage nurse sees an unsafe priority score but has no clear escalation control. Rather than applying the relevant control, it will make the review button visible but technically prevent staff from changing an AI recommendation. What can go wrong?

  1. People may be unable to correct known errors in time to avoid harm.
  2. The system may continue taking unsafe actions while teams debate a permanent correction.
  3. A user can receive restricted information or trigger an unauthorized operation.
  4. Sensitive inputs may be retained or reused beyond the organization's accepted boundaries.

Explanation

Correct Answer

A:

For the risk comparison, people may be unable to correct known errors in time to avoid harm. The shortcut overlooks the central safeguard: maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

C:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

D:

This consequence follows more directly from another shortcut: Send confidential records to a hosted model relying solely on the provider's marketing promise of privacy. The shortcut actually described would instead create this risk: People may be unable to correct known errors in time to avoid harm.

Question 730

The governance team needs both a decision and a way to verify it for this case: A hospital triage nurse sees an unsafe priority score but has no clear escalation control. Which pair is best?

  1. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  3. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  4. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.

Explanation

Correct Answer

B:

For the action and verification, maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Audit overridden cases for decision context, reviewer authority and follow-through. Together, they implement and test the intended safeguard. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate

Incorrect Answers

A:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

C:

This pair supports another governance intervention: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

D:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. The supporting evidence is: Audit overridden cases for decision context, reviewer authority and follow-through.

Question 731

An AI project team faces this situation: A municipal caseworker copies a chatbot explanation into a formal decision without checking sources. What should the team do first to manage the central issue?

  1. Use region- and role-aware controls when the permitted AI use differs across operating locations.
  2. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  3. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  4. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Explanation

Correct Answer

C:

Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The practical acceptance evidence is: use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Incorrect Answers

A:

This would help address a separate governance task: Use region- and role-aware controls when the permitted AI use differs across operating locations. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

B:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

D:

This would help address a separate governance task: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

Question 732

The organization chooses to teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty for this situation: A municipal caseworker copies a chatbot explanation into a formal decision without checking sources. Which rationale best supports that choice?

  1. A model's own wording cannot substitute for external authorization checks on data and actions.
  2. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  3. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  4. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Explanation

Correct Answer

D:

Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. That is why the case calls for this action: teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

Incorrect Answers

A:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

B:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

C:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Question 733

A municipal caseworker copies a chatbot explanation into a formal decision without checking sources. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  2. Test users with different roles against protected documents and disallowed tool operations.
  3. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  4. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Explanation

Correct Answer

A:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Incorrect Answers

B:

The alternative—test users with different roles against protected documents and disallowed tool operations—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

C:

The alternative—check logs and business processes for unapproved use cases, and require a new assessment when scope changes—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

D:

The alternative—test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Question 734

A team considers this AI use: A municipal caseworker copies a chatbot explanation into a formal decision without checking sources. Rather than applying the relevant control, it will treat access to an AI tool as proof that employees know how to use it safely. What can go wrong?

  1. A user can receive restricted information or trigger an unauthorized operation.
  2. Staff can act on fabricated answers or expose sensitive information through normal workflows.
  3. Unassessed harms may arise from new decision contexts and affected groups.
  4. Connectivity and data-handling requirements may make a technically capable model unusable.

Explanation

Correct Answer

B:

Staff can act on fabricated answers or expose sensitive information through normal workflows. The shortcut overlooks the central safeguard: teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

C:

This consequence follows more directly from another shortcut: Treat approval for inventory forecasting as approval to score individual employees' productivity. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

D:

This consequence follows more directly from another shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

Question 735

A municipal caseworker copies a chatbot explanation into a formal decision without checking sources. Which action and evidence pair most directly addresses the situation?

  1. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  2. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  3. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  4. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

C:

Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Use realistic work exercises to confirm users recognize unreliable content and apply the required check. Together, they implement and test the intended safeguard. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Incorrect Answers

A:

This pair supports another governance intervention: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

B:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

D:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Question 736

Consider the following proposed AI use: A sales team pastes confidential client contracts into a drafting assistant without understanding data rules. Which response best addresses the governance gap?

  1. Monitor quality, safety, latency and feedback against meaningful production thresholds.
  2. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  3. Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow.
  4. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

Explanation

Correct Answer

D:

For the decision, teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The practical acceptance evidence is: use realistic work exercises to confirm users recognize unreliable content and apply the required check

Incorrect Answers

A:

This would help address a separate governance task: Monitor quality, safety, latency and feedback against meaningful production thresholds. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

B:

This would help address a separate governance task: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

C:

This would help address a separate governance task: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For this case, however, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. The more direct step is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.

Question 737

A sales team pastes confidential client contracts into a drafting assistant without understanding data rules. Which explanation best connects the risk to the proposed governance action?

  1. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  2. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  3. Governance needs a response path for unsafe outcomes as well as conventional outages.
  4. Hosting architecture changes how data flows, how quickly decisions arrive and who can control system operations.

Explanation

Correct Answer

A:

For the technical explanation, effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output. That is why the case calls for this action: teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty

Incorrect Answers

B:

For the technical explanation, that explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

C:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

D:

That explains why a team might assess cloud, on-premises and edge deployment against latency, data residency, control and resource constraints, but it does not justify the decision required by the case. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

Question 738

An AI governance reviewer examines this situation: A sales team pastes confidential client contracts into a drafting assistant without understanding data rules. Which acceptance check is most decisive?

  1. Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  3. Test users with different roles against protected documents and disallowed tool operations.
  4. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Explanation

Correct Answer

B:

For the verification, use realistic work exercises to confirm users recognize unreliable content and apply the required check. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

Incorrect Answers

A:

The alternative—run a response exercise and verify incident logging, containment actions and post-incident review—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.

C:

For the verification, the alternative—test users with different roles against protected documents and disallowed tool operations—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

D:

For the verification, the alternative—test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks—would not resolve the stated issue. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

Question 739

A sales team pastes confidential client contracts into a drafting assistant without understanding data rules. The team instead proposes: Treat access to an AI tool as proof that employees know how to use it safely. Which failure should the reviewer expect most directly?

  1. Unsafe behavior may continue without a responsible containment decision.
  2. People may be unable to correct known errors in time to avoid harm.
  3. Staff can act on fabricated answers or expose sensitive information through normal workflows.
  4. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Explanation

Correct Answer

C:

For the risk comparison, staff can act on fabricated answers or expose sensitive information through normal workflows. The shortcut overlooks the central safeguard: teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

B:

This consequence follows more directly from another shortcut: Make the review button visible but technically prevent staff from changing an AI recommendation. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

D:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: Staff can act on fabricated answers or expose sensitive information through normal workflows.

Question 740

A sales team pastes confidential client contracts into a drafting assistant without understanding data rules. Which action and evidence pair most directly addresses the situation?

  1. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  2. Action: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. Evidence: Test users with different roles against protected documents and disallowed tool operations.
  3. Action: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. Evidence: Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  4. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Explanation

Correct Answer

D:

For the action and verification, teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Use realistic work exercises to confirm users recognize unreliable content and apply the required check. Together, they implement and test the intended safeguard. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check

B:

This pair supports another governance intervention: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

C:

This pair supports another governance intervention: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For the actual case, the required action is: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. The supporting evidence is: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.

Question 741

An education chatbot begins recommending dangerous behavior and the support desk has no escalation owner. Which step would provide the most targeted control for the problem described?

  1. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.
  2. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  3. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  4. Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model.

Explanation

Correct Answer

A:

Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Governance needs a response path for unsafe outcomes as well as conventional outages. The practical acceptance evidence is: run a response exercise and verify incident logging, containment actions and post-incident review.

Incorrect Answers

B:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

C:

This would help address a separate governance task: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

D:

This would help address a separate governance task: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

Question 742

An education chatbot begins recommending dangerous behavior and the support desk has no escalation owner. Which explanation best connects the risk to the proposed governance action?

  1. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  2. Governance needs a response path for unsafe outcomes as well as conventional outages.
  3. Human oversight requires a real opportunity to intervene before or after a high-impact automated recommendation, as appropriate.
  4. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Explanation

Correct Answer

B:

Governance needs a response path for unsafe outcomes as well as conventional outages. That is why the case calls for this action: define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

Incorrect Answers

A:

That explains why a team might plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

C:

That explains why a team might maintain usable override and escalation mechanisms for consequential AI-assisted decisions, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

D:

That explains why a team might define evidence access, incident notification and audit cooperation in third-party AI agreements, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

Question 743

An education chatbot begins recommending dangerous behavior and the support desk has no escalation owner. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  2. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  3. Run a response exercise and verify incident logging, containment actions and post-incident review.
  4. Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

C:

Run a response exercise and verify incident logging, containment actions and post-incident review. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance needs a response path for unsafe outcomes as well as conventional outages.

Incorrect Answers

A:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

B:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

D:

Run representative task trials measuring utility, error types, explainability and required operating controls would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

Question 744

For this case—An education chatbot begins recommending dangerous behavior and the support desk has no escalation owner—staff choose the shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. Which consequence best explains why this is inadequate?

  1. The system may perform restricted activities in places where different rules apply.
  2. People may be unable to correct known errors in time to avoid harm.
  3. Investigations and regulatory responses may be delayed by missing evidence rights.
  4. Unsafe behavior may continue without a responsible containment decision.

Explanation

Correct Answer

D:

Unsafe behavior may continue without a responsible containment decision. The shortcut overlooks the central safeguard: define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Governance needs a response path for unsafe outcomes as well as conventional outages.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision.

B:

This consequence follows more directly from another shortcut: Make the review button visible but technically prevent staff from changing an AI recommendation. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision.

C:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision.

Question 745

An education chatbot begins recommending dangerous behavior and the support desk has no escalation owner. Which action and evidence pair most directly addresses the situation?

  1. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  3. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  4. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

A:

Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Run a response exercise and verify incident logging, containment actions and post-incident review. Together, they implement and test the intended safeguard. Governance needs a response path for unsafe outcomes as well as conventional outages.

Incorrect Answers

B:

This pair supports another governance intervention: Monitor quality, safety, latency and feedback against meaningful production thresholds. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review.

C:

This pair supports another governance intervention: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review.

D:

This pair supports another governance intervention: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review.

Question 746

Consider the following proposed AI use: A customer-support model exposes fragments of private tickets during a live session. Which response best addresses the governance gap?

  1. Use region- and role-aware controls when the permitted AI use differs across operating locations.
  2. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.
  3. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.
  4. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.

Explanation

Correct Answer

B:

For the decision, define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Governance needs a response path for unsafe outcomes as well as conventional outages. The practical acceptance evidence is: run a response exercise and verify incident logging, containment actions and post-incident review

Incorrect Answers

A:

This would help address a separate governance task: Use region- and role-aware controls when the permitted AI use differs across operating locations. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

C:

This would help address a separate governance task: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

D:

This would help address a separate governance task: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For this case, however, governance needs a response path for unsafe outcomes as well as conventional outages. The more direct step is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.

Question 747

A team reviewing the following situation asks for the most relevant governance principle: A customer-support model exposes fragments of private tickets during a live session. Which principle best explains the required control?

  1. A model's own wording cannot substitute for external authorization checks on data and actions.
  2. Legal permission and local risk can depend on the place, population, processing and deployment context.
  3. Governance needs a response path for unsafe outcomes as well as conventional outages.
  4. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Explanation

Correct Answer

C:

For the technical explanation, governance needs a response path for unsafe outcomes as well as conventional outages. That is why the case calls for this action: define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents

Incorrect Answers

A:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

B:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

D:

That explains why a team might assess actual affected users, decisions, data and harms in the intended deployment environment, but it does not justify the decision required by the case. Governance needs a response path for unsafe outcomes as well as conventional outages.

Question 748

A customer-support model exposes fragments of private tickets during a live session. Which recorded test or observation most directly demonstrates an effective response?

  1. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  2. Test users with different roles against protected documents and disallowed tool operations.
  3. Document a fit assessment with real latency, connectivity, privacy and maintenance tests.
  4. Run a response exercise and verify incident logging, containment actions and post-incident review.

Explanation

Correct Answer

D:

For the verification, run a response exercise and verify incident logging, containment actions and post-incident review. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Governance needs a response path for unsafe outcomes as well as conventional outages

Incorrect Answers

A:

Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

B:

Test users with different roles against protected documents and disallowed tool operations would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

C:

Document a fit assessment with real latency, connectivity, privacy and maintenance tests would test a different outcome. The directly relevant evidence is to run a response exercise and verify incident logging, containment actions and post-incident review.

Question 749

A customer-support model exposes fragments of private tickets during a live session. The team instead proposes: Leave a harmful AI assistant running until the next quarterly governance meeting. Which failure should the reviewer expect most directly?

  1. Unsafe behavior may continue without a responsible containment decision.
  2. The system may perform restricted activities in places where different rules apply.
  3. The system may continue taking unsafe actions while teams debate a permanent correction.
  4. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Explanation

Correct Answer

A:

For the risk comparison, unsafe behavior may continue without a responsible containment decision. The shortcut overlooks the central safeguard: define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Governance needs a response path for unsafe outcomes as well as conventional outages

Incorrect Answers

B:

For the risk comparison, this consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision

C:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision.

D:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: Unsafe behavior may continue without a responsible containment decision.

Question 750

The governance team needs both a decision and a way to verify it for this case: A customer-support model exposes fragments of private tickets during a live session. Which pair is best?

  1. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  2. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  3. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  4. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

B:

For the action and verification, define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Run a response exercise and verify incident logging, containment actions and post-incident review. Together, they implement and test the intended safeguard. Governance needs a response path for unsafe outcomes as well as conventional outages

Incorrect Answers

A:

For the action and verification, this pair supports another governance intervention: Monitor quality, safety, latency and feedback against meaningful production thresholds. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review

C:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review.

D:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. The supporting evidence is: Run a response exercise and verify incident logging, containment actions and post-incident review.

Question 751

Consider the following proposed AI use: A warehouse model built for demand planning is reused to rank individual employee performance. Which response best addresses the governance gap?

  1. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  2. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  3. Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.
  4. Assess actual affected users, decisions, data and harms in the intended deployment environment.

Explanation

Correct Answer

C:

Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. A safe assessment for one use case does not establish suitability for a materially different downstream decision. The practical acceptance evidence is: check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Incorrect Answers

A:

The alternative—maintain usable override and escalation mechanisms for consequential AI-assisted decisions—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

B:

The alternative—apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

D:

The alternative—assess actual affected users, decisions, data and harms in the intended deployment environment—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

Question 752

A warehouse model built for demand planning is reused to rank individual employee performance. Why is the recommended governance approach appropriate?

  1. A model's own wording cannot substitute for external authorization checks on data and actions.
  2. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  3. Benchmark results obtained under another population, language or workflow may not transfer to local use.
  4. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Explanation

Correct Answer

D:

A safe assessment for one use case does not establish suitability for a materially different downstream decision. That is why the case calls for this action: review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

Incorrect Answers

A:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

B:

The alternative—deployment governance continues until systems are deactivated and their data, dependencies and users are addressed—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

C:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Question 753

A warehouse model built for demand planning is reused to rank individual employee performance. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  2. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  3. Run a response exercise and verify incident logging, containment actions and post-incident review.
  4. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.

Explanation

Correct Answer

A:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Incorrect Answers

B:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

C:

Run a response exercise and verify incident logging, containment actions and post-incident review would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

D:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Question 754

A warehouse model built for demand planning is reused to rank individual employee performance. The team instead proposes: Treat approval for inventory forecasting as approval to score individual employees' productivity. Which failure should the reviewer expect most directly?

  1. Stakeholders may be misled and unable to take appropriate protective steps.
  2. Unassessed harms may arise from new decision contexts and affected groups.
  3. Staff can act on fabricated answers or expose sensitive information through normal workflows.
  4. Unnecessary complexity, cost and uncontrolled outputs may appear where a narrower model would suffice.

Explanation

Correct Answer

B:

Unassessed harms may arise from new decision contexts and affected groups. The shortcut overlooks the central safeguard: review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups.

C:

This consequence follows more directly from another shortcut: Treat access to an AI tool as proof that employees know how to use it safely. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups.

D:

This consequence follows more directly from another shortcut: Adopt the largest generative model for every task without testing simpler alternatives. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups.

Question 755

A warehouse model built for demand planning is reused to rank individual employee performance. Which action and evidence pair most directly addresses the situation?

  1. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  2. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  3. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  4. Action: Assess actual affected users, decisions, data and harms in the intended deployment environment. Evidence: Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

C:

Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes. Together, they implement and test the intended safeguard. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Incorrect Answers

A:

This pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

B:

This pair supports another governance intervention: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

D:

This pair supports another governance intervention: Assess actual affected users, decisions, data and harms in the intended deployment environment. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Question 756

A public library's room booking assistant is repurposed to evaluate patrons' personal conduct. Which step would provide the most targeted control for the problem described?

  1. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  4. Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

Explanation

Correct Answer

D:

For the decision, review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. A safe assessment for one use case does not establish suitability for a materially different downstream decision. The practical acceptance evidence is: check logs and business processes for unapproved use cases, and require a new assessment when scope changes

Incorrect Answers

A:

The alternative—plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

B:

The alternative—teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

C:

The alternative—specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy—is not the most direct response to this case. The matching action is to review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.

Question 757

A public library's room booking assistant is repurposed to evaluate patrons' personal conduct. Which explanation best connects the risk to the proposed governance action?

  1. A safe assessment for one use case does not establish suitability for a materially different downstream decision.
  2. Legal permission and local risk can depend on the place, population, processing and deployment context.
  3. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  4. A deploying organization cannot respond adequately to failures if it has no access to necessary supplier evidence.

Explanation

Correct Answer

A:

For the technical explanation, a safe assessment for one use case does not establish suitability for a materially different downstream decision. That is why the case calls for this action: review whether an AI system is being repurposed for decisions, people or data beyond its approved scope

Incorrect Answers

B:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

C:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

D:

That explains why a team might define evidence access, incident notification and audit cooperation in third-party AI agreements, but it does not justify the decision required by the case. A safe assessment for one use case does not establish suitability for a materially different downstream decision.

Question 758

A public library's room booking assistant is repurposed to evaluate patrons' personal conduct. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  2. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  3. Audit overridden cases for decision context, reviewer authority and follow-through.
  4. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

B:

For the verification, check logs and business processes for unapproved use cases, and require a new assessment when scope changes. This verifies the necessary safeguard rather than merely showing that an AI project was completed. A safe assessment for one use case does not establish suitability for a materially different downstream decision

Incorrect Answers

A:

Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

C:

Audit overridden cases for decision context, reviewer authority and follow-through would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

D:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Question 759

A team considers this AI use: A public library's room booking assistant is repurposed to evaluate patrons' personal conduct. Rather than applying the relevant control, it will treat approval for inventory forecasting as approval to score individual employees' productivity. What can go wrong?

  1. Stakeholders may be misled and unable to take appropriate protective steps.
  2. A user can receive restricted information or trigger an unauthorized operation.
  3. Unassessed harms may arise from new decision contexts and affected groups.
  4. Answers can become stale and lack traceable sources after the source policy changes.

Explanation

Correct Answer

C:

For the risk comparison, unassessed harms may arise from new decision contexts and affected groups. The shortcut overlooks the central safeguard: review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. A safe assessment for one use case does not establish suitability for a materially different downstream decision

Incorrect Answers

A:

For the risk comparison, this consequence follows more directly from another shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups

B:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups.

D:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Unassessed harms may arise from new decision contexts and affected groups.

Question 760

Choose the intervention and validation that fit this business problem: A public library's room booking assistant is repurposed to evaluate patrons' personal conduct.

  1. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  3. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.
  4. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Explanation

Correct Answer

D:

For the action and verification, review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes. Together, they implement and test the intended safeguard. A safe assessment for one use case does not establish suitability for a materially different downstream decision

Incorrect Answers

A:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

B:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

C:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. The supporting evidence is: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.

Question 761

A transit provider launches an advisory chatbot without telling passengers about its limitations. Which step would provide the most targeted control for the problem described?

  1. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.
  2. Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.
  3. Maintain usable override and escalation mechanisms for consequential AI-assisted decisions.
  4. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.

Explanation

Correct Answer

A:

Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Trust and accountability depend on timely, factual communication of material changes and significant issues. The practical acceptance evidence is: test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Incorrect Answers

B:

This would help address a separate governance task: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For this case, however, trust and accountability depend on timely, factual communication of material changes and significant issues. The more direct step is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

C:

This would help address a separate governance task: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For this case, however, trust and accountability depend on timely, factual communication of material changes and significant issues. The more direct step is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

D:

This would help address a separate governance task: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For this case, however, trust and accountability depend on timely, factual communication of material changes and significant issues. The more direct step is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

Question 762

A transit provider launches an advisory chatbot without telling passengers about its limitations. Why is the recommended governance approach appropriate?

  1. Legal permission and local risk can depend on the place, population, processing and deployment context.
  2. Trust and accountability depend on timely, factual communication of material changes and significant issues.
  3. Governance needs a response path for unsafe outcomes as well as conventional outages.
  4. Natural-language generation and structured classification solve different problems and need different evaluations.

Explanation

Correct Answer

B:

Trust and accountability depend on timely, factual communication of material changes and significant issues. That is why the case calls for this action: develop an accurate incident and capability communication plan for users, customers and external stakeholders.

Incorrect Answers

A:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues.

C:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues.

D:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues.

Question 763

A transit provider launches an advisory chatbot without telling passengers about its limitations. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Audit overridden cases for decision context, reviewer authority and follow-through.
  3. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  4. Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks.

Explanation

Correct Answer

C:

Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Trust and accountability depend on timely, factual communication of material changes and significant issues.

Incorrect Answers

A:

Run a response exercise and verify incident logging, containment actions and post-incident review would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

B:

Audit overridden cases for decision context, reviewer authority and follow-through would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

D:

Test unauthorized tool calls, approval requirements, audit logs and safe termination of multi-step tasks would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Question 764

For this case—A transit provider launches an advisory chatbot without telling passengers about its limitations—staff choose the shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. Which consequence best explains why this is inadequate?

  1. A user can receive restricted information or trigger an unauthorized operation.
  2. Unsafe behavior may continue without a responsible containment decision.
  3. Answers can become stale and lack traceable sources after the source policy changes.
  4. Stakeholders may be misled and unable to take appropriate protective steps.

Explanation

Correct Answer

D:

Stakeholders may be misled and unable to take appropriate protective steps. The shortcut overlooks the central safeguard: develop an accurate incident and capability communication plan for users, customers and external stakeholders. Trust and accountability depend on timely, factual communication of material changes and significant issues.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Rely on the assistant's instruction to respect confidentiality while giving it unrestricted data access. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

B:

This consequence follows more directly from another shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

C:

This consequence follows more directly from another shortcut: Fine-tune once on a policy manual and assume future policy revisions will appear automatically in model weights. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

Question 765

A transit provider launches an advisory chatbot without telling passengers about its limitations. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  2. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  3. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  4. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Explanation

Correct Answer

A:

Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence. Together, they implement and test the intended safeguard. Trust and accountability depend on timely, factual communication of material changes and significant issues.

Incorrect Answers

B:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

C:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

D:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Question 766

A utility must notify customers after an AI error affected outage priority information. Which governance action most directly addresses the need in this situation?

  1. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  2. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.
  3. Monitor quality, safety, latency and feedback against meaningful production thresholds.
  4. Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit.

Explanation

Correct Answer

B:

For the decision, develop an accurate incident and capability communication plan for users, customers and external stakeholders. Trust and accountability depend on timely, factual communication of material changes and significant issues. The practical acceptance evidence is: test a disclosure scenario for message approval, accurate limitations, contacts and update cadence

Incorrect Answers

A:

The alternative—plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life—is not the most direct response to this case. The matching action is to develop an accurate incident and capability communication plan for users, customers and external stakeholders.

C:

This would help address a separate governance task: Monitor quality, safety, latency and feedback against meaningful production thresholds. For this case, however, trust and accountability depend on timely, factual communication of material changes and significant issues. The more direct step is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

D:

This would help address a separate governance task: Test grounded retrieval and controlled fine-tuning as separate methods for improving a language model's domain fit. For this case, however, trust and accountability depend on timely, factual communication of material changes and significant issues. The more direct step is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders.

Question 767

A utility must notify customers after an AI error affected outage priority information. Why is the recommended governance approach appropriate?

  1. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  2. Governance needs a response path for unsafe outcomes as well as conventional outages.
  3. Trust and accountability depend on timely, factual communication of material changes and significant issues.
  4. A generic supplier assessment may not describe the deployment organization's population or operating conditions.

Explanation

Correct Answer

C:

For the technical explanation, trust and accountability depend on timely, factual communication of material changes and significant issues. That is why the case calls for this action: develop an accurate incident and capability communication plan for users, customers and external stakeholders

Incorrect Answers

A:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues.

B:

For the technical explanation, that explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues

D:

That explains why a team might assess actual affected users, decisions, data and harms in the intended deployment environment, but it does not justify the decision required by the case. Trust and accountability depend on timely, factual communication of material changes and significant issues.

Question 768

The organization has implemented a control for this case: A utility must notify customers after an AI error affected outage priority information. Which result would best validate it?

  1. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  2. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  3. Evaluate equivalent examples on accuracy, latency, operating expense and security requirements.
  4. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Explanation

Correct Answer

D:

For the verification, test a disclosure scenario for message approval, accurate limitations, contacts and update cadence. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Trust and accountability depend on timely, factual communication of material changes and significant issues

Incorrect Answers

A:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

B:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

C:

Evaluate equivalent examples on accuracy, latency, operating expense and security requirements would test a different outcome. The directly relevant evidence is to test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Question 769

For this case—A utility must notify customers after an AI error affected outage priority information—staff choose the shortcut: Publish broad assurances that the model is always accurate while omitting known operational limits. Which consequence best explains why this is inadequate?

  1. Stakeholders may be misled and unable to take appropriate protective steps.
  2. The system may continue taking unsafe actions while teams debate a permanent correction.
  3. Shadow use, retained data or broken downstream decisions can persist after the announced retirement.
  4. Investigations and regulatory responses may be delayed by missing evidence rights.

Explanation

Correct Answer

A:

For the risk comparison, stakeholders may be misled and unable to take appropriate protective steps. The shortcut overlooks the central safeguard: develop an accurate incident and capability communication plan for users, customers and external stakeholders. Trust and accountability depend on timely, factual communication of material changes and significant issues

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

C:

This consequence follows more directly from another shortcut: Switch off a legacy model without checking whether dependent workflows still call it or retain data. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

D:

This consequence follows more directly from another shortcut: Assume the supplier will voluntarily share all incident logs despite no agreed process. The shortcut actually described would instead create this risk: Stakeholders may be misled and unable to take appropriate protective steps.

Question 770

Choose the intervention and validation that fit this business problem: A utility must notify customers after an AI error affected outage priority information.

  1. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  2. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  3. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  4. Action: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. Evidence: Compare supplier claims with local performance, bias and robustness tests using representative samples.

Explanation

Correct Answer

B:

For the action and verification, develop an accurate incident and capability communication plan for users, customers and external stakeholders. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence. Together, they implement and test the intended safeguard. Trust and accountability depend on timely, factual communication of material changes and significant issues

Incorrect Answers

A:

This pair supports another governance intervention: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

C:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

D:

This pair supports another governance intervention: Run deployment-representative tests instead of accepting supplier performance claims as sufficient proof. For the actual case, the required action is: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. The supporting evidence is: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.

Question 771

An AI project team faces this situation: A purchasing agent starts issuing unauthorized orders and no one knows how to halt its tools. What should the team do first to manage the central issue?

  1. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.
  4. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Explanation

Correct Answer

C:

Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The practical acceptance evidence is: exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Incorrect Answers

A:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

B:

This would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

D:

This would help address a separate governance task: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Question 772

A purchasing agent starts issuing unauthorized orders and no one knows how to halt its tools. Why is the recommended governance approach appropriate?

  1. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  2. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  3. An open-weight model or free download may still carry licensing, attribution, usage or distribution restrictions.
  4. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Explanation

Correct Answer

D:

An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. That is why the case calls for this action: implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Incorrect Answers

A:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

B:

That explains why a team might plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

C:

That explains why a team might verify model, dataset and output licensing conditions before integrating or redistributing AI components, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Question 773

A purchasing agent starts issuing unauthorized orders and no one knows how to halt its tools. Which recorded test or observation most directly demonstrates an effective response?

  1. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  2. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  3. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches.

Explanation

Correct Answer

A:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Incorrect Answers

B:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

C:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

D:

Compare grounded-answer accuracy, citation quality, maintenance effort and data governance across approaches would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Question 774

A team considers this AI use: A purchasing agent starts issuing unauthorized orders and no one knows how to halt its tools. Rather than applying the relevant control, it will assume a harmful autonomous workflow can always be fixed by editing its next prompt. What can go wrong?

  1. Unassessed harms may arise from new decision contexts and affected groups.
  2. The system may continue taking unsafe actions while teams debate a permanent correction.
  3. People may be unable to correct known errors in time to avoid harm.
  4. An agent may make unapproved purchases, record changes or external communications.

Explanation

Correct Answer

B:

The system may continue taking unsafe actions while teams debate a permanent correction. The shortcut overlooks the central safeguard: implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat approval for inventory forecasting as approval to score individual employees' productivity. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction.

C:

This consequence follows more directly from another shortcut: Make the review button visible but technically prevent staff from changing an AI recommendation. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction.

D:

This consequence follows more directly from another shortcut: Give an autonomous assistant broad write access because it handled informational questions well. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction.

Question 775

The governance team needs both a decision and a way to verify it for this case: A purchasing agent starts issuing unauthorized orders and no one knows how to halt its tools. Which pair is best?

  1. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  2. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  3. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  4. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Explanation

Correct Answer

C:

Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode. Together, they implement and test the intended safeguard. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Incorrect Answers

A:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

B:

This pair supports another governance intervention: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

D:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Question 776

Consider the following proposed AI use: A facility-control model issues unsafe scheduling commands and lacks a manual fallback. Which response best addresses the governance gap?

  1. Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents.
  2. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  3. Compare smaller and larger models on task performance, inference cost, energy use and maintainability.
  4. Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Explanation

Correct Answer

D:

For the decision, implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The practical acceptance evidence is: exercise the shutdown path and verify dependent workflows revert to a defined safe mode

Incorrect Answers

A:

This would help address a separate governance task: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

B:

For the decision, this would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback

C:

This would help address a separate governance task: Compare smaller and larger models on task performance, inference cost, energy use and maintainability. For this case, however, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. The more direct step is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback.

Question 777

A facility-control model issues unsafe scheduling commands and lacks a manual fallback. Why is the recommended governance approach appropriate?

  1. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.
  2. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  3. A model's own wording cannot substitute for external authorization checks on data and actions.
  4. Even a technically sound model can be unsafe when staff cannot interpret, contest or intervene in its outputs.

Explanation

Correct Answer

A:

For the technical explanation, an operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately. That is why the case calls for this action: implement and test a way to suspend unsafe AI functionality or switch to a safe fallback

Incorrect Answers

B:

For the technical explanation, that explains why a team might plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately

C:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

D:

That explains why a team might evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow, but it does not justify the decision required by the case. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.

Question 778

A facility-control model issues unsafe scheduling commands and lacks a manual fallback. Which recorded test or observation most directly demonstrates an effective response?

  1. Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  3. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  4. Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

B:

For the verification, exercise the shutdown path and verify dependent workflows revert to a defined safe mode. This verifies the necessary safeguard rather than merely showing that an AI project was completed. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately

Incorrect Answers

A:

Run a response exercise and verify incident logging, containment actions and post-incident review would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

C:

Inspect retirement records for disabled access, archive obligations, migration checks and owner approval would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

D:

Run representative task trials measuring utility, error types, explainability and required operating controls would test a different outcome. The directly relevant evidence is to exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Question 779

The organization has this problem: A facility-control model issues unsafe scheduling commands and lacks a manual fallback. It takes a shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. What is the principal risk?

  1. The system may perform restricted activities in places where different rules apply.
  2. Unassessed harms may arise from new decision contexts and affected groups.
  3. The system may continue taking unsafe actions while teams debate a permanent correction.
  4. The system may fail on the very cases the deploying organization must handle.

Explanation

Correct Answer

C:

For the risk comparison, the system may continue taking unsafe actions while teams debate a permanent correction. The shortcut overlooks the central safeguard: implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction.

B:

For the risk comparison, this consequence follows more directly from another shortcut: Treat approval for inventory forecasting as approval to score individual employees' productivity. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction

D:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: The system may continue taking unsafe actions while teams debate a permanent correction.

Question 780

A facility-control model issues unsafe scheduling commands and lacks a manual fallback. Which action and evidence pair most directly addresses the situation?

  1. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  2. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  3. Action: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. Evidence: Observe simulated decisions with trained users and record comprehension, override and escalation results.
  4. Action: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Evidence: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Explanation

Correct Answer

D:

For the action and verification, implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode. Together, they implement and test the intended safeguard. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately

Incorrect Answers

A:

This pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

B:

For the action and verification, this pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode

C:

This pair supports another governance intervention: Evaluate whether operators have training, capacity and escalation routes before activating an AI-supported workflow. For the actual case, the required action is: Implement and test a way to suspend unsafe AI functionality or switch to a safe fallback. The supporting evidence is: Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.

Question 781

A global employer enables candidate voice analysis in all offices without local legal checks. Which governance action most directly addresses the need in this situation?

  1. Use region- and role-aware controls when the permitted AI use differs across operating locations.
  2. Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope.
  3. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  4. Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors.

Explanation

Correct Answer

A:

Use region- and role-aware controls when the permitted AI use differs across operating locations. Legal permission and local risk can depend on the place, population, processing and deployment context. The practical acceptance evidence is: test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Incorrect Answers

B:

The alternative—review whether an AI system is being repurposed for decisions, people or data beyond its approved scope—would not resolve the stated issue. Legal permission and local risk can depend on the place, population, processing and deployment context.

C:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, legal permission and local risk can depend on the place, population, processing and deployment context. The more direct step is: Use region- and role-aware controls when the permitted AI use differs across operating locations.

D:

This would help address a separate governance task: Review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors. For this case, however, legal permission and local risk can depend on the place, population, processing and deployment context. The more direct step is: Use region- and role-aware controls when the permitted AI use differs across operating locations.

Question 782

The organization chooses to use region- and role-aware controls when the permitted AI use differs across operating locations for this situation: A global employer enables candidate voice analysis in all offices without local legal checks. Which rationale best supports that choice?

  1. Governance needs a response path for unsafe outcomes as well as conventional outages.
  2. Legal permission and local risk can depend on the place, population, processing and deployment context.
  3. A model's own wording cannot substitute for external authorization checks on data and actions.
  4. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Explanation

Correct Answer

B:

Legal permission and local risk can depend on the place, population, processing and deployment context. That is why the case calls for this action: use region- and role-aware controls when the permitted AI use differs across operating locations.

Incorrect Answers

A:

That explains why a team might define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

C:

That explains why a team might apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

D:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

Question 783

A global employer enables candidate voice analysis in all offices without local legal checks. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  2. Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  3. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  4. Trace each included model and dataset license to the planned product's commercial and distribution use.

Explanation

Correct Answer

C:

Test restricted-region access, user notices and fail-closed behavior against approved policy decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Legal permission and local risk can depend on the place, population, processing and deployment context.

Incorrect Answers

A:

Inspect retirement records for disabled access, archive obligations, migration checks and owner approval would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

B:

Check logs and business processes for unapproved use cases, and require a new assessment when scope changes would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

D:

Trace each included model and dataset license to the planned product's commercial and distribution use would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Question 784

A global employer enables candidate voice analysis in all offices without local legal checks. The team instead proposes: Deploy the same biometric feature worldwide because one region approved it. Which failure should the reviewer expect most directly?

  1. Unassessed harms may arise from new decision contexts and affected groups.
  2. Model harms and wrong decisions can rise while infrastructure remains healthy.
  3. The organization can underbudget safety, data and operations obligations it now owns.
  4. The system may perform restricted activities in places where different rules apply.

Explanation

Correct Answer

D:

The system may perform restricted activities in places where different rules apply. The shortcut overlooks the central safeguard: use region- and role-aware controls when the permitted AI use differs across operating locations. Legal permission and local risk can depend on the place, population, processing and deployment context.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Treat approval for inventory forecasting as approval to score individual employees' productivity. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

B:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

C:

This consequence follows more directly from another shortcut: Assume an in-house model removes vendor risk and therefore needs fewer governance controls. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

Question 785

The governance team needs both a decision and a way to verify it for this case: A global employer enables candidate voice analysis in all offices without local legal checks. Which pair is best?

  1. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  2. Action: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. Evidence: Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  3. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  4. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

A:

Use region- and role-aware controls when the permitted AI use differs across operating locations. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions. Together, they implement and test the intended safeguard. Legal permission and local risk can depend on the place, population, processing and deployment context.

Incorrect Answers

B:

This pair supports another governance intervention: Develop an accurate incident and capability communication plan for users, customers and external stakeholders. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

C:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

D:

This pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Question 786

A multinational retailer activates an identity-matching system across markets with different rules. Which governance action most directly addresses the need in this situation?

  1. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  2. Use region- and role-aware controls when the permitted AI use differs across operating locations.
  3. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  4. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.

Explanation

Correct Answer

B:

For the decision, use region- and role-aware controls when the permitted AI use differs across operating locations. Legal permission and local risk can depend on the place, population, processing and deployment context. The practical acceptance evidence is: test restricted-region access, user notices and fail-closed behavior against approved policy decisions

Incorrect Answers

A:

This would help address a separate governance task: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For this case, however, legal permission and local risk can depend on the place, population, processing and deployment context. The more direct step is: Use region- and role-aware controls when the permitted AI use differs across operating locations.

C:

For the decision, this would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, legal permission and local risk can depend on the place, population, processing and deployment context. The more direct step is: Use region- and role-aware controls when the permitted AI use differs across operating locations

D:

This would help address a separate governance task: Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy. For this case, however, legal permission and local risk can depend on the place, population, processing and deployment context. The more direct step is: Use region- and role-aware controls when the permitted AI use differs across operating locations.

Question 787

A multinational retailer activates an identity-matching system across markets with different rules. Which explanation best connects the risk to the proposed governance action?

  1. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  2. Trust and accountability depend on timely, factual communication of material changes and significant issues.
  3. Legal permission and local risk can depend on the place, population, processing and deployment context.
  4. Contractual restrictions determine whether supplied data may be retained or used for model improvement.

Explanation

Correct Answer

C:

For the technical explanation, legal permission and local risk can depend on the place, population, processing and deployment context. That is why the case calls for this action: use region- and role-aware controls when the permitted AI use differs across operating locations

Incorrect Answers

A:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

B:

That explains why a team might develop an accurate incident and capability communication plan for users, customers and external stakeholders, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

D:

That explains why a team might review AI vendor clauses governing prompts, records, retention, training reuse and subprocessors, but it does not justify the decision required by the case. Legal permission and local risk can depend on the place, population, processing and deployment context.

Question 788

An AI governance reviewer examines this situation: A multinational retailer activates an identity-matching system across markets with different rules. Which acceptance check is most decisive?

  1. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  2. Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence.
  3. Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise.
  4. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Explanation

Correct Answer

D:

For the verification, test restricted-region access, user notices and fail-closed behavior against approved policy decisions. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Legal permission and local risk can depend on the place, population, processing and deployment context

Incorrect Answers

A:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

B:

Test a disclosure scenario for message approval, accurate limitations, contacts and update cadence would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

C:

Test contractual notice deadlines, documentation access and incident-contact procedures in a tabletop exercise would test a different outcome. The directly relevant evidence is to test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

Question 789

The organization has this problem: A multinational retailer activates an identity-matching system across markets with different rules. It takes a shortcut: Deploy the same biometric feature worldwide because one region approved it. What is the principal risk?

  1. The system may perform restricted activities in places where different rules apply.
  2. Unsafe behavior may continue without a responsible containment decision.
  3. People may be unable to correct known errors in time to avoid harm.
  4. The system may fail on the very cases the deploying organization must handle.

Explanation

Correct Answer

A:

For the risk comparison, the system may perform restricted activities in places where different rules apply. The shortcut overlooks the central safeguard: use region- and role-aware controls when the permitted AI use differs across operating locations. Legal permission and local risk can depend on the place, population, processing and deployment context

Incorrect Answers

B:

This consequence follows more directly from another shortcut: Leave a harmful AI assistant running until the next quarterly governance meeting. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

C:

This consequence follows more directly from another shortcut: Make the review button visible but technically prevent staff from changing an AI recommendation. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

D:

This consequence follows more directly from another shortcut: Approve a model because the supplier reports excellent global benchmark accuracy without local testing. The shortcut actually described would instead create this risk: The system may perform restricted activities in places where different rules apply.

Question 790

A multinational retailer activates an identity-matching system across markets with different rules. Which plan includes both the appropriate control and a meaningful check?

  1. Action: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. Evidence: Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  2. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  3. Action: Monitor quality, safety, latency and feedback against meaningful production thresholds. Evidence: Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  4. Action: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. Evidence: Run representative task trials measuring utility, error types, explainability and required operating controls.

Explanation

Correct Answer

B:

For the action and verification, use region- and role-aware controls when the permitted AI use differs across operating locations. Test restricted-region access, user notices and fail-closed behavior against approved policy decisions. Together, they implement and test the intended safeguard. Legal permission and local risk can depend on the place, population, processing and deployment context

Incorrect Answers

A:

This pair supports another governance intervention: Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

C:

This pair supports another governance intervention: Monitor quality, safety, latency and feedback against meaningful production thresholds. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.

D:

For the action and verification, this pair supports another governance intervention: Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes. For the actual case, the required action is: Use region- and role-aware controls when the permitted AI use differs across operating locations. The supporting evidence is: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions

Question 791

Consider the following proposed AI use: A nonprofit replaces an old donor-scoring API but leaves scheduled jobs using its credentials. Which response best addresses the governance gap?

  1. Monitor quality, safety, latency and feedback against meaningful production thresholds.
  2. Develop an accurate incident and capability communication plan for users, customers and external stakeholders.
  3. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.
  4. Compare classic predictive models with generative approaches against the defined problem and acceptable failure modes.

Explanation

Correct Answer

C:

Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. The practical acceptance evidence is: inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Incorrect Answers

A:

This would help address a separate governance task: Monitor quality, safety, latency and feedback against meaningful production thresholds. For this case, however, deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. The more direct step is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

B:

The alternative—develop an accurate incident and capability communication plan for users, customers and external stakeholders—is not the most direct response to this case. The matching action is to plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

D:

The alternative—compare classic predictive models with generative approaches against the defined problem and acceptable failure modes—is not the most direct response to this case. The matching action is to plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

Question 792

A nonprofit replaces an old donor-scoring API but leaves scheduled jobs using its credentials. Which explanation best connects the risk to the proposed governance action?

  1. Effective deployment depends on calibrated reliance rather than unquestioning acceptance of fluent AI output.
  2. An operational kill switch reduces continued harm when unacceptable behavior cannot be corrected immediately.
  3. Natural-language generation and structured classification solve different problems and need different evaluations.
  4. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Explanation

Correct Answer

D:

Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. That is why the case calls for this action: plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

Incorrect Answers

A:

That explains why a team might teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

B:

That explains why a team might implement and test a way to suspend unsafe AI functionality or switch to a safe fallback, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

C:

That explains why a team might compare classic predictive models with generative approaches against the defined problem and acceptable failure modes, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Question 793

The organization has implemented a control for this case: A nonprofit replaces an old donor-scoring API but leaves scheduled jobs using its credentials. Which result would best validate it?

  1. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  2. Test users with different roles against protected documents and disallowed tool operations.
  3. Exercise the shutdown path and verify dependent workflows revert to a defined safe mode.
  4. Compare local impact assessment findings with vendor claims and obtain risk-owner signoff.

Explanation

Correct Answer

A:

Inspect retirement records for disabled access, archive obligations, migration checks and owner approval. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Incorrect Answers

B:

Test users with different roles against protected documents and disallowed tool operations would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

C:

Exercise the shutdown path and verify dependent workflows revert to a defined safe mode would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

D:

Compare local impact assessment findings with vendor claims and obtain risk-owner signoff would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Question 794

For this case—A nonprofit replaces an old donor-scoring API but leaves scheduled jobs using its credentials—staff choose the shortcut: Switch off a legacy model without checking whether dependent workflows still call it or retain data. Which consequence best explains why this is inadequate?

  1. The system may continue taking unsafe actions while teams debate a permanent correction.
  2. Shadow use, retained data or broken downstream decisions can persist after the announced retirement.
  3. People may be unable to correct known errors in time to avoid harm.
  4. The deployment may violate license conditions or require unexpected changes.

Explanation

Correct Answer

B:

Shadow use, retained data or broken downstream decisions can persist after the announced retirement. The shortcut overlooks the central safeguard: plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Assume a harmful autonomous workflow can always be fixed by editing its next prompt. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

C:

This consequence follows more directly from another shortcut: Make the review button visible but technically prevent staff from changing an AI recommendation. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

D:

This consequence follows more directly from another shortcut: Describe every openly downloadable model as unrestricted public-domain software. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

Question 795

A nonprofit replaces an old donor-scoring API but leaves scheduled jobs using its credentials. Which action and evidence pair most directly addresses the situation?

  1. Action: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. Evidence: Run a response exercise and verify incident logging, containment actions and post-incident review.
  2. Action: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. Evidence: Audit overridden cases for decision context, reviewer authority and follow-through.
  3. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  4. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.

Explanation

Correct Answer

C:

Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval. Together, they implement and test the intended safeguard. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Incorrect Answers

A:

This pair supports another governance intervention: Define how to stop harm, preserve evidence, notify owners and recover from AI deployment incidents. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

B:

This pair supports another governance intervention: Maintain usable override and escalation mechanisms for consequential AI-assisted decisions. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

D:

This pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Question 796

A city agency decommissions a permit assistant without archiving complaints and approval evidence. Which governance action most directly addresses the need in this situation?

  1. Teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty.
  2. Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation.
  3. Specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy.
  4. Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

Explanation

Correct Answer

D:

For the decision, plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. The practical acceptance evidence is: inspect retirement records for disabled access, archive obligations, migration checks and owner approval

Incorrect Answers

A:

The alternative—teach operational users when to validate outputs, protect sensitive inputs and escalate uncertainty—is not the most direct response to this case. The matching action is to plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

B:

This would help address a separate governance task: Apply least-privilege permissions to inputs, retrieved documents, tools and outputs during operation. For this case, however, deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. The more direct step is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

C:

The alternative—specify agent goals, approved tools, permissions, human checkpoints and action limits before enabling autonomy—is not the most direct response to this case. The matching action is to plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life.

Question 797

The organization chooses to plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life for this situation: A city agency decommissions a permit assistant without archiving complaints and approval evidence. Which rationale best supports that choice?

  1. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.
  2. Legal permission and local risk can depend on the place, population, processing and deployment context.
  3. Availability alone does not reveal harmful answers, performance drift or rising decision error rates.
  4. Benchmark results obtained under another population, language or workflow may not transfer to local use.

Explanation

Correct Answer

A:

For the technical explanation, deployment governance continues until systems are deactivated and their data, dependencies and users are addressed. That is why the case calls for this action: plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life

Incorrect Answers

B:

That explains why a team might use region- and role-aware controls when the permitted AI use differs across operating locations, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

C:

That explains why a team might monitor quality, safety, latency and feedback against meaningful production thresholds, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

D:

That explains why a team might run deployment-representative tests instead of accepting supplier performance claims as sufficient proof, but it does not justify the decision required by the case. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed.

Question 798

A city agency decommissions a permit assistant without archiving complaints and approval evidence. After taking the recommended governance action, which evidence most directly verifies that the risk has been addressed?

  1. Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches.
  2. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.
  3. Use realistic work exercises to confirm users recognize unreliable content and apply the required check.
  4. Observe simulated decisions with trained users and record comprehension, override and escalation results.

Explanation

Correct Answer

B:

For the verification, inspect retirement records for disabled access, archive obligations, migration checks and owner approval. This verifies the necessary safeguard rather than merely showing that an AI project was completed. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed

Incorrect Answers

A:

Inspect live dashboards, alert tests and escalation tickets for quality and safety threshold breaches would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

C:

Use realistic work exercises to confirm users recognize unreliable content and apply the required check would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

D:

Observe simulated decisions with trained users and record comprehension, override and escalation results would test a different outcome. The directly relevant evidence is to inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Question 799

A city agency decommissions a permit assistant without archiving complaints and approval evidence. The team instead proposes: Switch off a legacy model without checking whether dependent workflows still call it or retain data. Which failure should the reviewer expect most directly?

  1. Model harms and wrong decisions can rise while infrastructure remains healthy.
  2. The system may perform restricted activities in places where different rules apply.
  3. Shadow use, retained data or broken downstream decisions can persist after the announced retirement.
  4. Connectivity and data-handling requirements may make a technically capable model unusable.

Explanation

Correct Answer

C:

For the risk comparison, shadow use, retained data or broken downstream decisions can persist after the announced retirement. The shortcut overlooks the central safeguard: plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed

Incorrect Answers

A:

This consequence follows more directly from another shortcut: Watch only server uptime as evidence that an AI-assisted service is performing acceptably. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

B:

This consequence follows more directly from another shortcut: Deploy the same biometric feature worldwide because one region approved it. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

D:

This consequence follows more directly from another shortcut: Send every real-time sensitive inference to a remote cloud service without examining operating conditions. The shortcut actually described would instead create this risk: Shadow use, retained data or broken downstream decisions can persist after the announced retirement.

Question 800

Choose the intervention and validation that fit this business problem: A city agency decommissions a permit assistant without archiving complaints and approval evidence.

  1. Action: Use region- and role-aware controls when the permitted AI use differs across operating locations. Evidence: Test restricted-region access, user notices and fail-closed behavior against approved policy decisions.
  2. Action: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. Evidence: Check logs and business processes for unapproved use cases, and require a new assessment when scope changes.
  3. Action: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. Evidence: Review an end-to-end ownership and cost matrix covering provenance, testing, monitoring and potential liabilities.
  4. Action: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Evidence: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

Explanation

Correct Answer

D:

For the action and verification, plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. Inspect retirement records for disabled access, archive obligations, migration checks and owner approval. Together, they implement and test the intended safeguard. Deployment governance continues until systems are deactivated and their data, dependencies and users are addressed

Incorrect Answers

A:

This pair supports another governance intervention: Use region- and role-aware controls when the permitted AI use differs across operating locations. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

B:

This pair supports another governance intervention: Review whether an AI system is being repurposed for decisions, people or data beyond its approved scope. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval.

C:

For the action and verification, this pair supports another governance intervention: Evaluate additional lifecycle, security, governance and legal responsibilities before choosing an in-house foundation model. For the actual case, the required action is: Plan review of feedback, model replacement, archival evidence and safe retirement at the end of AI service life. The supporting evidence is: Inspect retirement records for disabled access, archive obligations, migration checks and owner approval