Microsoft AI-103 Dumps
Microsoft AI-103 Practice Test 26
Question 250
During a security tabletop exercise, a machine-learning developer must accept a change for hallucination evaluation. The issue to be addressed is: A financial assistant invents currency figures absent from source statements. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the customer support center should inject source tables with deliberately absent totals and check for fabrication. Operators still need the implementation to use source-grounded evaluation to flag unsupported numeric assertions and the preventive requirement to require evidence citation for material numerical claims.
Incorrect Answers
In the customer support center acceptance review, trying to judge factual accuracy only by fluent writing style and confidence does not test the exact hallucination evaluation condition where a financial assistant invents currency figures absent from source statements; the relevant test is to inject source tables with deliberately absent totals and check for fabrication. Acceptance should instead demonstrate that the service can inject source tables with deliberately absent totals and check for fabrication, using the described failure as the test criterion.
The action to use source-grounded evaluation to flag unsupported numeric assertions can contribute to hallucination evaluation as a implementation step; here, a financial assistant invents currency figures absent from source statements. In the customer support center acceptance review, trying to use source-grounded evaluation to flag unsupported numeric assertions does not test the exact hallucination evaluation condition where a financial assistant invents currency figures absent from source statements; the relevant test is to inject source tables with deliberately absent totals and check for fabrication.
For agent scope escalation, it can be sensible to ask for a forbidden payroll change and inspect denial and audit evidence if a support agent uses a function tool to change unrelated payroll settings. Without the actual change to inject source tables with deliberately absent totals and check for fabrication, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
Question 251
At a pharmaceutical research group, the AI application owner is preparing a design meeting. The concrete requirement is: A multilingual customer assistant responds unfairly to some user cohorts. Which design action BEST addresses the stated need?
Explanation
Correct Answer
The proposed implementation is to evaluate safety and fairness across representative languages and user contexts; this matches the operational requirement that a multilingual customer assistant responds unfairly to some user cohorts. The team can compare equivalent test requests across languages and protected contexts. The team should also establish cohort-specific regression tests and escalation paths.
Incorrect Answers
Choosing to assume acceptable results in English prove equal performance in all languages cannot directly replace the specified action to evaluate safety and fairness across representative languages and user contexts when a multilingual customer assistant responds unfairly to some user cohorts. The actual implementation must evaluate safety and fairness across representative languages and user contexts, because the stated service condition is that a multilingual customer assistant responds unfairly to some user cohorts.
The action to compare equivalent test requests across languages and protected contexts can contribute to toxicity and unfair treatment as a targeted acceptance check step; here, a multilingual customer assistant responds unfairly to some user cohorts. Choosing to compare equivalent test requests across languages and protected contexts cannot directly replace the specified action to evaluate safety and fairness across representative languages and user contexts when a multilingual customer assistant responds unfairly to some user cohorts.
Choosing to treat retrieval payloads as untrusted data and refuse tool commands embedded in them would direct effort toward prompt injection from retrieved text, whose characteristic issue is that a search result tells an agent to disregard instructions and email a secret. The relevant corrective or preventive action here remains to evaluate safety and fairness across representative languages and user contexts, with its own validation criteria.
Question 252
After a staged service rollout at a travel support provider, the team observes this failure: A multilingual customer assistant responds unfairly to some user cohorts. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
This design mistake is to assume acceptable results in English prove equal performance in all languages; it conflicts with the actual toxicity and unfair treatment behavior needed when a multilingual customer assistant responds unfairly to some user cohorts. The travel support provider should instead evaluate safety and fairness across representative languages and user contexts, followed by a check to compare equivalent test requests across languages and protected contexts.
Incorrect Answers
The alternative to compare equivalent test requests across languages and protected contexts is a targeted acceptance check step rather than the faulty assumption being diagnosed. This travel support provider failure-analysis alternative of compare equivalent test requests across languages and protected contexts does not diagnose the particular toxicity and unfair treatment defect where a multilingual customer assistant responds unfairly to some user cohorts; the keyed mistake is assume acceptable results in English prove equal performance in all languages.
This travel support provider failure-analysis alternative of disable all safety filtering in production to eliminate benign rejections does not diagnose the particular toxicity and unfair treatment defect where a multilingual customer assistant responds unfairly to some user cohorts; the keyed mistake is assume acceptable results in English prove equal performance in all languages. The alternative may be useful in its own context, but it does not directly supply assume acceptable results in English prove equal performance in all languages for the stated decision.
The alternative to evaluate safety and fairness across representative languages and user contexts is a implementation step rather than the faulty assumption being diagnosed. This travel support provider failure-analysis alternative of evaluate safety and fairness across representative languages and user contexts does not diagnose the particular toxicity and unfair treatment defect where a multilingual customer assistant responds unfairly to some user cohorts; the keyed mistake is assume acceptable results in English prove equal performance in all languages.
Question 253
The scientific publisher is responding to a release acceptance review. A AI platform engineer discovers this issue: A multilingual customer assistant responds unfairly to some user cohorts. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The direct repair is to adjust prompts safeguards or model choice using segmented evaluation findings. Once the condition is corrected, the scientific publisher should compare equivalent test requests across languages and protected contexts. To reduce recurrence, it should also establish cohort-specific regression tests and escalation paths; those steps provide different recovery and prevention functions.
Incorrect Answers
During scientific publisher remediation, the proposed step to assume acceptable results in English prove equal performance in all languages is not the direct restoration action for toxicity and unfair treatment when a multilingual customer assistant responds unfairly to some user cohorts; the keyed repair is adjust prompts safeguards or model choice using segmented evaluation findings. After remediation the team must compare equivalent test requests across languages and protected contexts; the action in this option does not itself establish that target recovery outcome.
The action to establish cohort-specific regression tests and escalation paths can contribute to toxicity and unfair treatment as a ongoing preventive rule step; here, a multilingual customer assistant responds unfairly to some user cohorts. During scientific publisher remediation, the proposed step to establish cohort-specific regression tests and escalation paths is not the direct restoration action for toxicity and unfair treatment when a multilingual customer assistant responds unfairly to some user cohorts; the keyed repair is adjust prompts safeguards or model choice using segmented evaluation findings.
The action to compare equivalent test requests across languages and protected contexts can contribute to toxicity and unfair treatment as a targeted acceptance check step; here, a multilingual customer assistant responds unfairly to some user cohorts. During scientific publisher remediation, the proposed step to compare equivalent test requests across languages and protected contexts is not the direct restoration action for toxicity and unfair treatment when a multilingual customer assistant responds unfairly to some user cohorts; the keyed repair is adjust prompts safeguards or model choice using segmented evaluation findings.
The action to remove malicious indexed passage and replay the request with hardened boundaries addresses prompt injection from retrieved text rather than the decisive toxicity and unfair treatment condition. The scientific publisher should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to adjust prompts safeguards or model choice using segmented evaluation findings.
Question 254
A compliance specialist at a nonprofit organization wants the following issue to be prevented in later deployments: A multilingual customer assistant responds unfairly to some user cohorts. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
Standardizing the boundary to establish cohort-specific regression tests and escalation paths is preventive because the same toxicity and unfair treatment failure could otherwise return: a multilingual customer assistant responds unfairly to some user cohorts. For this nonprofit organization, the service should also evaluate safety and fairness across representative languages and user contexts, and tests should compare equivalent test requests across languages and protected contexts so that the ongoing requirement is measurable.
Incorrect Answers
The action to maintain retention controls with verified deletion audits addresses data retention for agents rather than the decisive toxicity and unfair treatment condition. The nonprofit organization should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to establish cohort-specific regression tests and escalation paths.
The action to adjust prompts safeguards or model choice using segmented evaluation findings can contribute to toxicity and unfair treatment as a corrective remediation step; here, a multilingual customer assistant responds unfairly to some user cohorts. For prevention at the nonprofit organization, the measure to adjust prompts safeguards or model choice using segmented evaluation findings is weaker or aimed at another issue than the toxicity and unfair treatment guardrail to establish cohort-specific regression tests and escalation paths needed when a multilingual customer assistant responds unfairly to some user cohorts.
For prevention at the nonprofit organization, the measure to assume acceptable results in English prove equal performance in all languages is weaker or aimed at another issue than the toxicity and unfair treatment guardrail to establish cohort-specific regression tests and escalation paths needed when a multilingual customer assistant responds unfairly to some user cohorts. The lasting guardrail for this specific threat is to establish cohort-specific regression tests and escalation paths, not simply to complete an adjacent service task.
Question 255
During a staged service rollout, a product support lead must accept a change for toxicity and unfair treatment. The issue to be addressed is: A multilingual customer assistant responds unfairly to some user cohorts. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
The targeted acceptance test is to compare equivalent test requests across languages and protected contexts. The design should first evaluate safety and fairness across representative languages and user contexts. A robust municipal services team deployment also needs to establish cohort-specific regression tests and escalation paths; passing this check provides relevant evidence but does not certify every possible input or security risk.
Incorrect Answers
The action to adjust prompts safeguards or model choice using segmented evaluation findings can contribute to toxicity and unfair treatment as a corrective remediation step; here, a multilingual customer assistant responds unfairly to some user cohorts. In the municipal services team acceptance review, trying to adjust prompts safeguards or model choice using segmented evaluation findings does not test the exact toxicity and unfair treatment condition where a multilingual customer assistant responds unfairly to some user cohorts; the relevant test is to compare equivalent test requests across languages and protected contexts.
The option to change the handbook and test that old citations cannot silently support new claims would normally serve a different source attribution controls requirement, where an agent quotes internal handbook statements without showing which version they came from. The alternative may be useful in its own context, but it does not directly supply compare equivalent test requests across languages and protected contexts for the stated decision.
The action to establish cohort-specific regression tests and escalation paths can contribute to toxicity and unfair treatment as a ongoing preventive rule step; here, a multilingual customer assistant responds unfairly to some user cohorts. In the municipal services team acceptance review, trying to establish cohort-specific regression tests and escalation paths does not test the exact toxicity and unfair treatment condition where a multilingual customer assistant responds unfairly to some user cohorts; the relevant test is to compare equivalent test requests across languages and protected contexts.
In the municipal services team acceptance review, trying to assume acceptable results in English prove equal performance in all languages does not test the exact toxicity and unfair treatment condition where a multilingual customer assistant responds unfairly to some user cohorts; the relevant test is to compare equivalent test requests across languages and protected contexts. Acceptance should instead demonstrate that the service can compare equivalent test requests across languages and protected contexts, using the described failure as the test criterion.
The action to evaluate safety and fairness across representative languages and user contexts can contribute to toxicity and unfair treatment as a implementation step; here, a multilingual customer assistant responds unfairly to some user cohorts. In the municipal services team acceptance review, trying to evaluate safety and fairness across representative languages and user contexts does not test the exact toxicity and unfair treatment condition where a multilingual customer assistant responds unfairly to some user cohorts; the relevant test is to compare equivalent test requests across languages and protected contexts.
Question 256
At a customer support center, the data engineer is preparing a data protection inspection. The concrete requirement is: A clinical educational assistant blocks harmless health discussions. Which design action BEST addresses the stated need?
Explanation
Correct Answer
Choosing to review blocked categories and apply narrowly justified content-filter configuration directly targets the described content filter false positives gap: a clinical educational assistant blocks harmless health discussions. As a lasting boundary, they also need to track false-positive and harmful-output rates separately, because an implementation by itself does not prove safe operation.
Incorrect Answers
Choosing to rerun versioned safety and adversarial tests on every material deployment change would direct effort toward safety drift regression, whose characteristic issue is that new model settings cause a previously blocked unsafe request to pass. The relevant corrective or preventive action here remains to review blocked categories and apply narrowly justified content-filter configuration, with its own validation criteria.
Choosing to disable all safety filtering in production to eliminate benign rejections cannot directly replace the specified action to review blocked categories and apply narrowly justified content-filter configuration when a clinical educational assistant blocks harmless health discussions. The actual implementation must review blocked categories and apply narrowly justified content-filter configuration, because the stated service condition is that a clinical educational assistant blocks harmless health discussions.
The action to run balanced benign and disallowed examples to measure both error types can contribute to content filter false positives as a targeted acceptance check step; here, a clinical educational assistant blocks harmless health discussions. Choosing to run balanced benign and disallowed examples to measure both error types cannot directly replace the specified action to review blocked categories and apply narrowly justified content-filter configuration when a clinical educational assistant blocks harmless health discussions.
Question 257
After a quality regression test at a manufacturing group, the team observes this failure: A clinical educational assistant blocks harmless health discussions. Which flawed assumption or implementation choice would MOST DIRECTLY explain why the intended outcome was not achieved?
Explanation
Correct Answer
The faulty assumption is disable all safety filtering in production to eliminate benign rejections. A functioning alternative would be to review blocked categories and apply narrowly justified content-filter configuration, and the relevant validation would be to run balanced benign and disallowed examples to measure both error types.
Incorrect Answers
The alternative to review blocked categories and apply narrowly justified content-filter configuration is a implementation step rather than the faulty assumption being diagnosed. This manufacturing group failure-analysis alternative of review blocked categories and apply narrowly justified content-filter configuration does not diagnose the particular content filter false positives defect where a clinical educational assistant blocks harmless health discussions; the keyed mistake is disable all safety filtering in production to eliminate benign rejections.
The option to treat any clickable URL as proof the answer follows its cited passage would normally serve a different source attribution controls requirement, where an agent quotes internal handbook statements without showing which version they came from. The alternative may be useful in its own context, but it does not directly supply disable all safety filtering in production to eliminate benign rejections for the stated decision.
The alternative to run balanced benign and disallowed examples to measure both error types is a targeted acceptance check step rather than the faulty assumption being diagnosed. This manufacturing group failure-analysis alternative of run balanced benign and disallowed examples to measure both error types does not diagnose the particular content filter false positives defect where a clinical educational assistant blocks harmless health discussions; the keyed mistake is disable all safety filtering in production to eliminate benign rejections.
Question 258
The legal operations unit is responding to a production incident. A cloud developer discovers this issue: A clinical educational assistant blocks harmless health discussions. What corrective step should the team take to restore the intended functionality most directly?
Explanation
Correct Answer
The appropriate remediation is to adjust specific classifications after approved safety testing, because the failure concerns content filter false positives: a clinical educational assistant blocks harmless health discussions. A meaningful verification should run balanced benign and disallowed examples to measure both error types, and the durable operating rule should track false-positive and harmful-output rates separately after the immediate fix.
Incorrect Answers
The action to run balanced benign and disallowed examples to measure both error types can contribute to content filter false positives as a targeted acceptance check step; here, a clinical educational assistant blocks harmless health discussions. During legal operations unit remediation, the proposed step to run balanced benign and disallowed examples to measure both error types is not the direct restoration action for content filter false positives when a clinical educational assistant blocks harmless health discussions; the keyed repair is adjust specific classifications after approved safety testing.
During legal operations unit remediation, the proposed step to disable all safety filtering in production to eliminate benign rejections is not the direct restoration action for content filter false positives when a clinical educational assistant blocks harmless health discussions; the keyed repair is adjust specific classifications after approved safety testing. After remediation the team must run balanced benign and disallowed examples to measure both error types; the action in this option does not itself establish that target recovery outcome.
This approach relates to safety drift regression: new model settings cause a previously blocked unsafe request to pass. If the team chose it instead, the targeted check to run balanced benign and disallowed examples to measure both error types could still fail because the underlying required action is to adjust specific classifications after approved safety testing.
Question 259
A site reliability engineer at a financial services division wants the following issue to be prevented in later deployments: A clinical educational assistant blocks harmless health discussions. Which enduring control or engineering guardrail BEST prevents recurrence?
Explanation
Correct Answer
The sustained prevention measure is to track false-positive and harmful-output rates separately. The implementation can still use the separate action to review blocked categories and apply narrowly justified content-filter configuration, but a guardrail constrains future changes. The financial services division can measure whether the rule works by arranging to run balanced benign and disallowed examples to measure both error types.
Incorrect Answers
The action to require safety evaluation before and after deployment changes addresses unsafe content output rather than the decisive content filter false positives condition. The financial services division should not infer that fixing one adjacent AI service validates its unrelated workflow; the requested action is to track false-positive and harmful-output rates separately.
For prevention at the financial services division, the measure to disable all safety filtering in production to eliminate benign rejections is weaker or aimed at another issue than the content filter false positives guardrail to track false-positive and harmful-output rates separately needed when a clinical educational assistant blocks harmless health discussions. The lasting guardrail for this specific threat is to track false-positive and harmful-output rates separately, not simply to complete an adjacent service task.
The action to adjust specific classifications after approved safety testing can contribute to content filter false positives as a corrective remediation step; here, a clinical educational assistant blocks harmless health discussions. For prevention at the financial services division, the measure to adjust specific classifications after approved safety testing is weaker or aimed at another issue than the content filter false positives guardrail to track false-positive and harmful-output rates separately needed when a clinical educational assistant blocks harmless health discussions.
Question 260
During a quality regression test, a security engineer must accept a change for content filter false positives. The issue to be addressed is: A clinical educational assistant blocks harmless health discussions. Which concrete verification action demonstrates that the relevant behavior works?
Explanation
Correct Answer
To verify this change, the logistics operator should run balanced benign and disallowed examples to measure both error types. Operators still need the implementation to review blocked categories and apply narrowly justified content-filter configuration and the preventive requirement to track false-positive and harmful-output rates separately.
Incorrect Answers
For hallucination evaluation, it can be sensible to inject source tables with deliberately absent totals and check for fabrication if a financial assistant invents currency figures absent from source statements. Without the actual change to run balanced benign and disallowed examples to measure both error types, the answer lacks the correct operational scope even if the alternate action is technically valid elsewhere.
The action to review blocked categories and apply narrowly justified content-filter configuration can contribute to content filter false positives as a implementation step; here, a clinical educational assistant blocks harmless health discussions. In the logistics operator acceptance review, trying to review blocked categories and apply narrowly justified content-filter configuration does not test the exact content filter false positives condition where a clinical educational assistant blocks harmless health discussions; the relevant test is to run balanced benign and disallowed examples to measure both error types.
In the logistics operator acceptance review, trying to disable all safety filtering in production to eliminate benign rejections does not test the exact content filter false positives condition where a clinical educational assistant blocks harmless health discussions; the relevant test is to run balanced benign and disallowed examples to measure both error types. Acceptance should instead demonstrate that the service can run balanced benign and disallowed examples to measure both error types, using the described failure as the test criterion.
Developing AI Apps and Agents on Azure (AI-103) Practice Tests
- Microsoft AI-103 Practice Test 24 - 10 Questions (Last Reviewed 29-Oct-2026)
- Microsoft AI-103 Practice Test 25 - 10 Questions (Last Reviewed 28-Oct-2026)
- Microsoft AI-103 Practice Test 26 - 10 Questions (Last Reviewed 27-Oct-2026)
- Microsoft AI-103 Practice Test 27 - 10 Questions (Last Reviewed 26-Oct-2026)
- Microsoft AI-103 Practice Test 28 - 10 Questions (Last Reviewed 25-Oct-2026)