AWS Certified AI Practitioner AIF-C01 Practice Test 43
Question 420
A GenAI assistant must avoid repeating social security numbers found in user-provided text. Before approving a rollout, an architect suggests Use public CloudTrail events as a repository for raw sensitive prompts. Which mismatch should be documented for this precise design?
- Sampling does not reliably prevent information disclosure
- Unnecessary prompt logging can increase exposure
- Sensitive data handling needs explicit controls rather than model goodwill
- A polite request is not enforceable protection
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Use public CloudTrail events as a repository for raw sensitive prompts. Unnecessary prompt logging can increase exposure
Incorrect Answers
A:This limitation concerns Increase model temperature so identifiers become random; it does not diagnose the proposed Use public CloudTrail events as a repository for raw sensitive prompts approach.
C:Sensitive data handling needs explicit controls rather than model goodwill This explains the recommended design, but it is not the requested diagnosis of Use public CloudTrail events as a repository for raw sensitive prompts.
D:This limitation concerns Rely only on a natural-language request to keep secrets; it does not diagnose the proposed Use public CloudTrail events as a repository for raw sensitive prompts approach.
Question 421
A GenAI assistant must avoid repeating social security numbers found in user-provided text. Which option BEST meets the business requirement?
- Configure sensitive-information protections and minimize PII exposure
- Rely only on a natural-language request to keep secrets
- Use public CloudTrail events as a repository for raw sensitive prompts
- Increase model temperature so identifiers become random
Explanation
Correct Answer
A:Sensitive data handling needs explicit controls rather than model goodwill The specified scenario makes this the most appropriate approach.
Incorrect Answers
B:A polite request is not enforceable protection This prevents the approach from meeting the stated requirement.
C:Unnecessary prompt logging can increase exposure This prevents the approach from meeting the stated requirement.
D:Sampling does not reliably prevent information disclosure This prevents the approach from meeting the stated requirement.
Question 422
A regulated firm must document a model's intended use, limitations and risk rating. A pilot uses this approach instead: Use billing invoices as the complete model-risk record. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Weights alone do not describe governance decisions
- A storage path cannot capture evaluation and intended use
- Cost reports omit intended uses and technical limitations
- Model Cards capture model metadata, intended uses and risk considerations
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Use billing invoices as the complete model-risk record. Cost reports omit intended uses and technical limitations
Incorrect Answers
A:This limitation concerns Assume a saved model weight file documents model lineage and risk; it does not diagnose the proposed Use billing invoices as the complete model-risk record approach.
B:This limitation concerns Treat a dataset location as the entire model documentation; it does not diagnose the proposed Use billing invoices as the complete model-risk record approach.
D:Model Cards capture model metadata, intended uses and risk considerations This explains the recommended design, but it is not the requested diagnosis of Use billing invoices as the complete model-risk record.
Question 423
An analyst needs a model explanation that business users can understand without revealing training secrets. During a design review, someone proposes Use only neural network layer counts for every decision. Which technical objection is specifically relevant to this proposal?
- Architecture size does not explain a particular prediction
- Useful explainability describes factors affecting predictions in an appropriate form
- Exposing source records can violate privacy and still be confusing
- Charges do not establish why the model responded
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use only neural network layer counts for every decision. Architecture size does not explain a particular prediction
Incorrect Answers
B:Useful explainability describes factors affecting predictions in an appropriate form This explains the recommended design, but it is not the requested diagnosis of Use only neural network layer counts for every decision.
C:This limitation concerns Reveal private training rows as the explanation; it does not diagnose the proposed Use only neural network layer counts for every decision approach.
D:This limitation concerns Provide a billing receipt instead of decision rationale; it does not diagnose the proposed Use only neural network layer counts for every decision approach.
Question 424
A bank asks how to communicate performance limitations for a fraud model. Which option BEST meets the business requirement?
- Publish only the model's hidden weights
- Publish a current model card with intended uses, limitations and evaluated metrics
- Publish only a marketing statement about AI accuracy
- Provide only a service uptime chart as model quality disclosure
Explanation
Correct Answer
B:Transparent documentation lets users understand where output is reliable The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:Weights cannot serve as a plain-language limitation record This prevents the approach from meeting the stated requirement.
C:Marketing copy omits measured limitations and assumptions This prevents the approach from meeting the stated requirement.
D:Availability does not describe failure modes or bias This prevents the approach from meeting the stated requirement.
Question 425
A credit-screening model performs well overall but misses qualified applicants in one protected group. Before approving a rollout, an architect suggests Remove demographic evaluation fields before measuring disparities. Which mismatch should be documented for this precise design?
- A high average score can coexist with harmful group-level errors
- Without appropriate controlled subgroup information the disparity may remain hidden
- Size does not measure differential impact
- Aggregate metrics can hide disparities across populations
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Remove demographic evaluation fields before measuring disparities. Without appropriate controlled subgroup information the disparity may remain hidden
Incorrect Answers
A:This limitation concerns Use overall accuracy as the only fairness measure; it does not diagnose the proposed Remove demographic evaluation fields before measuring disparities approach.
C:This limitation concerns Judge fairness solely by model parameter count; it does not diagnose the proposed Remove demographic evaluation fields before measuring disparities approach.
D:Aggregate metrics can hide disparities across populations This explains the recommended design, but it is not the requested diagnosis of Remove demographic evaluation fields before measuring disparities.
Question 426
A GenAI assistant must avoid repeating social security numbers found in user-provided text. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- Negative test cases show protected identifiers are masked or rejected as intended
- A reviewed design record approves this competing approach: Use public CloudTrail events as a repository for raw sensitive prompts
- A system inventory lists this competing approach as active: Rely only on a natural-language request to keep secrets
- A component test invokes this alternate approach successfully: Increase model temperature so identifiers become random
Explanation
Correct Answer
A:Negative test cases show protected identifiers are masked or rejected as intended. This checks the business result rather than just a configured feature.
Incorrect Answers
B:Unnecessary prompt logging can increase exposure Design approval is not evidence that the alternative meets the actual requirement.
C:A polite request is not enforceable protection An active setting cannot independently prove the requested business outcome.
D:Sampling does not reliably prevent information disclosure A component invocation does not validate the necessary end-to-end outcome.
Question 427
A high-impact customer eligibility model has borderline cases needing a human decision. Before approving a rollout, an architect suggests Use lower latency as proof that human review is unnecessary. Which mismatch should be documented for this precise design?
- Toxicity filtering does not replace human eligibility judgment
- Faster inference is not evidence of acceptable decision risk
- Reviewers can inspect uncertain outputs before critical action
- Confidence alone may not meet policy for sensitive outcomes
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Use lower latency as proof that human review is unnecessary. Faster inference is not evidence of acceptable decision risk
Incorrect Answers
A:This limitation concerns Treat a guardrail toxicity filter as the final eligibility reviewer; it does not diagnose the proposed Use lower latency as proof that human review is unnecessary approach.
C:Reviewers can inspect uncertain outputs before critical action This explains the recommended design, but it is not the requested diagnosis of Use lower latency as proof that human review is unnecessary.
D:This limitation concerns Automatically approve every high-confidence output without defined controls; it does not diagnose the proposed Use lower latency as proof that human review is unnecessary approach.
Question 428
A deployed predictor starts receiving different input distributions from the data used for validation. Before approving a rollout, an architect suggests Measure only S3 storage growth as model drift. Which mismatch should be documented for this precise design?
- Deployment state says nothing about future distribution shift
- Uncontrolled retraining can introduce new risk
- Storage bytes do not diagnose feature distribution change
- Production input changes can degrade predictive behavior
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Measure only S3 storage growth as model drift. Storage bytes do not diagnose feature distribution change
Incorrect Answers
A:This limitation concerns Assume a successful deployment permanently guarantees accuracy; it does not diagnose the proposed Measure only S3 storage growth as model drift approach.
B:This limitation concerns Retrain continuously without tracking data or performance changes; it does not diagnose the proposed Measure only S3 storage growth as model drift approach.
D:Production input changes can degrade predictive behavior This explains the recommended design, but it is not the requested diagnosis of Measure only S3 storage growth as model drift.
Question 429
An external chat assistant sometimes returns insulting or hateful language. Which option BEST meets the business requirement?
- Use only a vector database index as the content safety filter
- Apply Amazon Bedrock Guardrails content filters with testing
- Rely only on a polite system persona without safety testing
- Change IAM billing permissions to remove toxic language
Explanation
Correct Answer
B:Guardrails can help filter unwanted model inputs and outputs The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:Retrieval search is not an output toxicity control This prevents the approach from meeting the stated requirement.
C:Prompt wording alone is not a complete moderation control This prevents the approach from meeting the stated requirement.
D:Billing authorization does not inspect generated text This prevents the approach from meeting the stated requirement.
Question 430
A high-impact customer eligibility model has borderline cases needing a human decision. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A reviewed design record approves this competing approach: Use lower latency as proof that human review is unnecessary
- A system inventory lists this competing approach as active: Automatically approve every high-confidence output without defined controls
- Recorded reviewer decisions resolve flagged cases before final action
- A component test invokes this alternate approach successfully: Treat a guardrail toxicity filter as the final eligibility reviewer
Explanation
Correct Answer
C:Recorded reviewer decisions resolve flagged cases before final action. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Faster inference is not evidence of acceptable decision risk Design approval is not evidence that the alternative meets the actual requirement.
B:Confidence alone may not meet policy for sensitive outcomes An active setting cannot independently prove the requested business outcome.
D:Toxicity filtering does not replace human eligibility judgment A component invocation does not validate the necessary end-to-end outcome.
