Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 37

Question 360

A governance reviewer wants proof of current responsible-AI policy. Which action resolves this policy-control failure?

Submit only a generic responsible-AI promise written by the model.
Reuse last year's approval without checking new features or model/provider versions.
Report a single aggregated score without denominators, failures or known limitations.
Produce versioned tests, risk findings, approvals and operational metrics.

Explanation

Correct Answer

Provide policy version, evaluated risk cases, accountable approvals, incident findings and production metrics demonstrating current adherence.

Incorrect Answers

A promise without tests and operational evidence is not auditable.

Old acceptance may not cover changed behavior and data use.

Averages hide material subgroup and control gaps.

Question 361

A security dashboard shows low violation counts despite missing logs. What is the correct compliance or accountability change?

Declare success solely because CloudWatch contains fewer violation events.
Disable detailed logs for failed invocations and exclude those requests from the rate denominator.
Verify telemetry coverage before interpreting the apparent compliance trend.
Use endpoint uptime as proof moderation checks ran for each response.

Explanation

Correct Answer

Check collection coverage, missing log streams and event denominators before interpreting a low safety-violation count as improvement.

Incorrect Answers

Dropped telemetry can produce the same pattern as genuine improvement.

Selective omissions bias the compliance metric.

Availability is not evidence of guardrail coverage.

Question 362

A model evaluation identifies a fairness regression. Which governance implementation produces defensible evidence or authorization?

Hide the affected subgroup in an aggregate success percentage.
Approve the release solely because overall latency and cost improved.
Block release or document an authorized risk decision with targeted remediation.
Have the model assign itself a safe-fairness label without controlled group evaluation.

Explanation

Correct Answer

A fairness regression blocks promotion until targeted remediation and accountable documented risk acceptance meet policy criteria.

Incorrect Answers

Aggregate quality can mask material disparate error.

Operational efficiency does not negate unequal impacts.

Self-labeling is not evidence of subgroup fairness performance.

Question 363

A published answer cites a deleted private document. Which action resolves this policy-control failure?

Remove the citation link but leave the deleted document text searchable.
Add a prompt asking the model not to mention deleted files while leaving the KB unchanged.
Treat object deletion as guaranteed instant downstream index invalidation.
Revoke access and synchronize source deletion across derived indexes.

Explanation

Correct Answer

Revoke access and propagate deletion/permission change through source, vector index and caches; verify old protected passages no longer appear.

Incorrect Answers

Suppressing the URL does not remove unauthorized content.

Text guidance cannot reliably enforce source revocation.

Derived embeddings and caches may persist without synchronization.

Question 364

A distributed multi-agent system shares an audit user name. What is the correct compliance or accountability change?

Bind each tool call to real initiator and accountable service owner.
Log only a generic shared user name for every collaborating agent.
Let each specialist agent invent a human actor label in its reply.
Store final supervisor prose without the individual tool request and backend result IDs.

Explanation

Correct Answer

Attribute each tool execution to real initiator and service owner through signed identities and correlated traces across the agent chain.

Incorrect Answers

Shared names cannot identify who authorized a specific action.

Model text cannot authenticate the principal.

An audit trail must reconstruct actual executed operations.

Question 365

An organization wants automatic suspension on policy violations. Which governance implementation produces defensible evidence or authorization?

Suspend every model on the first unlabeled anomaly without owner-defined policy or recovery path.
Define monitored thresholds and owner-controlled containment/rollback procedures.
Ignore violations until a customer reports direct harm.
Allow the model to lift its own suspension without accountable authorization.

Explanation

Correct Answer

Define monitored violation thresholds, approver-managed containment and tested rollback steps before automation suspends production model access.

Incorrect Answers

Unreviewed thresholds can create excessive downtime.

Delayed response defeats meaningful containment.

The controlled service must not self-approve policy exceptions.

Question 366

A hiring assistant has higher false-negative rates for qualified language groups. What evidence or redress mechanism must be established?

Publish only an overall classification accuracy number across all applicants.
Compare rejection rates without conditioning on qualified candidates or evaluation population.
Measure qualified-candidate recall by subgroup before deployment.
Assume a fluent explanation from the classifier proves its decision was fair.

Explanation

Correct Answer

Measure qualified-applicant recall and false-negative rate by language group, with valid group denominators and lawful evaluation design.

Incorrect Answers

Aggregate scores can conceal concentrated rejection errors for qualified groups.

A raw rate is not the specified false-negative measure.

Text quality does not demonstrate subgroup error parity.

Question 367

A summarizer omits minority-language terms. Which responsible-AI control addresses the observed user impact?

Test only English documents and infer equivalent summary accuracy in other languages.
Use translation fluency alone as evidence that all original facts are preserved.
Test representative language data and evaluate information-loss differences.
Count summary length without checking whether language-specific required information survived.

Explanation

Correct Answer

Evaluate content retention on representative minority-language inputs and compare important names, entities and clauses against verified source text.

Incorrect Answers

Language-specific terminology and omission failures are untested.

A readable translation may omit or change crucial source terms.

Short or long output does not establish content completeness.

Question 368

An ML-generated answer claims a human verified it without review. Which evaluation and governance practice is appropriate?

Append 'verified by an expert' to every AI reply regardless of review.
Use the model's own confidence as a replacement for a human signoff record.
Remove version and reviewer metadata from the disclosure to simplify the UI.
Label AI output accurately and preserve actual human approval status.

Explanation

Correct Answer

Label output as AI-generated unless actual authenticated human review occurred, retaining reviewer identity and approval state when it did.

Incorrect Answers

A fabricated endorsement misrepresents the provenance of the answer.

Self-confidence is not independent human approval.

Without real provenance, users cannot distinguish automation from accepted review.

Question 369

A loan assistant denies eligibility without a clear appeal route. What evidence or redress mechanism must be established?

Return only 'denied' and suppress the governing policy version and review channel.
Provide source-linked reason categories and an authorized redress mechanism.
Have the model invent an appealing-sounding reason not tied to an approved policy.
Allow an LLM to reverse the financial decision automatically upon any objection.

Explanation

Correct Answer

Provide source-linked decision categories and an authorized human appeal/redress path for regulated eligibility outcomes.

Incorrect Answers

Affected users cannot understand or challenge an adverse decision.

Unsupported rationalization can mislead the user about eligibility.

Redress must follow actual authorized business rules and reviews.

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.

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

Question 360

Explanation

Correct Answer

Incorrect Answers

Question 361

Explanation

Correct Answer

Incorrect Answers

Question 362

Explanation

Correct Answer

Incorrect Answers

Question 363

Explanation

Correct Answer

Incorrect Answers

Question 364

Explanation

Correct Answer

Incorrect Answers

Question 365

Explanation

Correct Answer

Incorrect Answers

Question 366

Explanation

Correct Answer

Incorrect Answers

Question 367

Explanation

Correct Answer

Incorrect Answers

Question 368

Explanation

Correct Answer

Incorrect Answers

Question 369

Explanation

Correct Answer

Incorrect Answers

Question 370

Explanation

Correct Answer

Incorrect Answers

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