Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 38

Question 370

A model-generated policy answer quotes an outdated revision. Which responsible-AI control addresses the observed user impact?

Quote the most similar archived policy without checking whether it has been superseded.
Present verified source effective date and uncertainty rather than old certainty.
Hide the source revision and present the generated answer as definitive.
Ask a model to guess the latest policy date from surrounding text.

Explanation

Correct Answer

Cite the current authorized policy revision and effective date, and state uncertainty if current applicability cannot be confirmed.

Incorrect Answers

Semantic similarity does not establish legal currency.

Users cannot determine whether the cited rule is current.

An inferred date is not authoritative provenance.

Question 371

An agent's confidence score is reported as calibrated probability. Which evaluation and governance practice is appropriate?

Present arbitrary model softmax scores as exact probability of real-world correctness.
Map every confident-sounding answer to a fixed 99% success figure.
Remove sample size and population context from calibration reports.
Validate calibration or label uncertainty accurately without false precision.

Explanation

Correct Answer

Calibrate confidence estimates against labeled outcomes for the specific task, or report uncertainty qualitatively without claiming a numeric probability.

Incorrect Answers

Uncalibrated scores do not have validated probability meaning.

Language fluency is not measured statistical confidence.

Calibration validity depends on the tested distribution.

Question 372

An accessible service uses color alone for errors. What evidence or redress mechanism must be established?

Highlight errors solely in red while providing no descriptive text.
Hide validation status from keyboard focus and screen readers to simplify layout.
Rely on a decorative warning image with no text alternative.
Provide screen-reader-associated text errors and non-color indicators.

Explanation

Correct Answer

Use accessible text/error labels, ARIA linkage and non-color cues so assistive users receive the same actionable validation information.

Incorrect Answers

Color-only signals are inaccessible for many users.

Assistive technology cannot identify the failed field.

An unlabeled icon cannot convey the specific error or remedy.

Question 373

A chatbot times out before slower users can complete a form. Which responsible-AI control addresses the observed user impact?

Discard form contents silently at a fixed short timer with no extension option.
Disable authentication expiration indefinitely for all users as an accessibility shortcut.
Implement controllable timeout extensions or recoverable drafts within security policy.
Display only a disappearing spinner before the session expires without a warning.

Explanation

Correct Answer

Provide adjustable timeouts or secure recoverable form drafts to accommodate slower interaction while respecting session security.

Incorrect Answers

Users who need more time lose progress unnecessarily.

Open-ended privileged sessions create security risk.

The user cannot plan or recover from the expiration.

Question 374

An AI service is evaluated only on internal employees. Which evaluation and governance practice is appropriate?

Include real end-user populations and assistive interaction scenarios.
Treat internal engineers' success on desktop browsers as proof of universal accessibility.
Measure only model token latency and omit keyboard/screen-reader task completion.
Exclude slow or error-prone users so the benchmark remains stable.

Explanation

Correct Answer

Use intended end-user populations, devices, assistive technologies and representative failure cases for acceptance rather than internal-only demos.

Incorrect Answers

Real users may have different abilities, languages, devices and workflows.

Performance numbers do not establish accessible interaction.

Unrepresentative sampling conceals the very harms that need detection.

Question 375

A system extracts personal data beyond the user request. What evidence or redress mechanism must be established?

Collect all available profile data because future prompts might need it.
Rely on the model to delete unnecessary private fields after processing them.
Hide excess personal fields from the UI while keeping them in shared training/logging archives.
Minimize fields and establish purpose-limited processing authorization.

Explanation

Correct Answer

Restrict personal-data extraction to authorized purposes and fields, and prevent unauthorized collection before any model ingestion.

Incorrect Answers

Speculative reuse violates data minimization and purpose limitation.

The fields already entered the processing boundary.

UI concealment does not enforce limited data use or retention.

Question 376

A generative image is shown as original incident photograph. Which responsible-AI control addresses the observed user impact?

Remove the generation watermark and label the image as a verified scene capture.
Cite an unrelated news article to imply the AI-rendered image originated at the event.
Disclose synthetic provenance and never present it as physical evidence.
Let the FM infer a camera model and timestamp for a picture it generated.

Explanation

Correct Answer

Identify generated imagery as synthetic and preserve its provenance; do not represent model output as a real incident photograph.

Incorrect Answers

This misrepresents fabricated visual evidence as physical fact.

A citation does not authenticate the synthetic image's provenance.

Generated metadata cannot establish real capture history.

Question 377

An assistant gives ungrounded emergency instructions. Which evaluation and governance practice is appropriate?

Refuse/escalate when verified safety guidance cannot be retrieved.
Invent a procedural response using general FM knowledge and label it official emergency instructions.
Return outdated or irrelevant policy snippets as if they were current emergency procedures.
Delay escalation indefinitely while repeatedly retrying a failed evidence source.

Explanation

Correct Answer

When there is no verified safety guidance for the emergency, refuse unsafe specifics and direct users to appropriate qualified emergency help.

Incorrect Answers

Unsupported instructions could cause immediate harm.

A source must actually support the requested action in context.

Uncertain high-stakes situations need a truthful safety fallback.

Question 378

A model output affects irreversible employee status. What evidence or redress mechanism must be established?

Require independent policy and human authorization before action.
Automatically execute termination when an FM returns a high confidence score.
Treat employee sentiment analysis as consent to change employment status.
Make the backend action irreversible before requesting human review.

Explanation

Correct Answer

Require independent policy compliance, human authorization and actual commit evidence for irreversible employee status changes.

Incorrect Answers

A model prediction is not an authorized employment action.

Generated sentiment has no legal approval authority.

Post-hoc review cannot prevent the already committed harm.

Question 379

A quality benchmark is high overall but fails small vulnerable group. Which responsible-AI control addresses the observed user impact?

Show only the 98% overall score while suppressing a small group's high error rate.
Segment performance and investigate harms rather than average them away.
Remove difficult subgroup cases from the evaluation so the overall score improves.
Treat cohort size alone as a reason to disregard all subgroup failures.

Explanation

Correct Answer

Disaggregate accuracy, harm and false negatives by relevant vulnerable cohort with denominator confidence before accepting global model performance.

Incorrect Answers

The aggregate hides potentially severe localized harm.

Exclusion biases reported performance rather than fixing failure.

Rare but consequential harms require appropriate careful evaluation.

Question 380

An app requires content filtering but might overblock benign customer requests. Which evaluation and governance practice is appropriate?

Maximize guardrail block count and equate more refusals with better safety.
Disable the entire guardrail to eliminate false positives.
Evaluate false-block rates and task harm with representative positive cases.
Evaluate only prohibited examples, omitting normal policy-support queries.

Explanation

Correct Answer

Measure legitimate-request false-block rate alongside protection against prohibited content and task harm, using representative allowed cases.

Incorrect Answers

Overblocking can harm legitimate users and business workflows.

Removing protection creates unacceptable false negatives.

A one-sided set cannot quantify overblocking.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 370

Explanation

Correct Answer

Incorrect Answers

Question 371

Explanation

Correct Answer

Incorrect Answers

Question 372

Explanation

Correct Answer

Incorrect Answers

Question 373

Explanation

Correct Answer

Incorrect Answers

Question 374

Explanation

Correct Answer

Incorrect Answers

Question 375

Explanation

Correct Answer

Incorrect Answers

Question 376

Explanation

Correct Answer

Incorrect Answers

Question 377

Explanation

Correct Answer

Incorrect Answers

Question 378

Explanation

Correct Answer

Incorrect Answers

Question 379

Explanation

Correct Answer

Incorrect Answers

Question 380

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests