Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 39

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.

Question 381

A customer complains of biased account classification. What evidence or redress mechanism must be established?

Assign human review owner and record corrections with relevant evidence.
Ask the model to declare its own classification unbiased without independent investigation.
Erase the complaint after a generated apology without reviewing the actual account record.
Automatically overturn all similar classifications without validating policy and affected records.

Explanation

Correct Answer

Assign an accountable human reviewer who checks classification evidence, records corrective action and follows an authorized appeal process.

Incorrect Answers

Self-assertion is not a correction or an appeal decision.

A polite message cannot repair a materially wrong classification.

Uncontrolled bulk changes can introduce new errors and harm.

Question 382

A team promises all GenAI recommendations are correct. Which responsible-AI control addresses the observed user impact?

Promise users 100% correct recommendations because the model achieved a strong benchmark.
Disclose evaluated limitations and unsupported cases in user documentation.
Suppress documented unsupported cases to present a more confident product description.
Use an issuer logo as implied endorsement of the application's answer quality.

Explanation

Correct Answer

Describe evaluated use cases, limitations, known failure modes and escalation paths; GenAI correctness cannot be guaranteed outside validated scope.

Incorrect Answers

A benchmark does not establish universal correctness.

Incomplete disclosure misleads users about reliance risks.

Hosting/certification does not certify every app-level recommendation.

Question 383

A response explains policy using inaccessible jargon. Which evaluation and governance practice is appropriate?

Remove all policy exceptions when simplifying the text for readability.
Repeat technical jargon verbatim while labeling the output accessible.
Invent a simple but inaccurate rule of thumb not supported by the approved source.
Provide reviewed plain-language explanation without dropping required conditions.

Explanation

Correct Answer

Use reviewed plain-language explanations that preserve essential eligibility, exception and appeal conditions with accurate source links.

Incorrect Answers

Simplification that changes the rule can mislead decisions.

Jargon without explanation remains hard to understand.

Ease of reading does not justify a false policy claim.

Question 384

A model revision changes behavior without informing operators. What evidence or redress mechanism must be established?

Change model aliases silently and provide only a green health dashboard.
Document release scope, limits and escalation responsibilities.
Discard previous model acceptance cases when a new version goes live.
Tell operators the model is unchanged because the UI branding is identical.

Explanation

Correct Answer

Give operators the precise model/prompt change log, new limitations, rollback and escalation owner so they can respond to regressions.

Incorrect Answers

Infrastructure health cannot explain unexpected behavior drift.

Historical test evidence is needed to diagnose changed outputs.

The model/version behavior can change independently of the interface.

Question 385

A deployment meets latency SLA but gives inaccurate guidance. Which responsible-AI control addresses the observed user impact?

Report grounded correctness and unsafe answer rate alongside operational uptime.
Report 99.9% uptime as proof the assistant's advice is correct.
Use token generation rate as the only indicator of policy compliance.
Stop evaluation after the endpoint responds successfully, without checking its content.

Explanation

Correct Answer

Track grounded factual correctness, harmful/unjustified response rates and evidence support alongside the ordinary latency/availability SLA.

Incorrect Answers

Availability does not evaluate answer truth.

Throughput measures performance, not supported claims.

A technically successful call may still produce incorrect guidance.

Question 386

A support app sends an entire manual for every short request. Which optimization balances actual cost and accepted task quality?

Always pass the full 500-page manual because a longer prompt is necessarily more accurate.
Retrieve only relevant source chunks and measure token reduction.
Cache the entire manual in every user's system prompt even when only one policy is requested.
Truncate the generation output only, leaving the oversized manual in each input.

Explanation

Correct Answer

Retrieve only relevant authorized manual passages and compare actual token savings with supported answer quality against a full-manual baseline.

Incorrect Answers

Unneeded tokens increase cost and context overflow risk without guaranteeing better grounding.

Repeated large context undermines efficient retrieval and user-specific authorization.

Output caps do not reduce redundant input token spend.

Question 387

A caller requests a one-line status but model emits long essays. Which resource-efficiency decision fits the workload?

Set suitable output limits and measure completion quality.
Remove all content limits because the model should decide how much detail the user needs.
Cut the input's safety and transaction constraints solely to save tokens.
Select a cheaper model without checking whether it follows the one-line status contract.

Explanation

Correct Answer

Cap output tokens and shape concise answers to the status task, then evaluate whether the required status details remain correct.

Incorrect Answers

Unbounded essays increase latency and cost for a one-line task.

Removing essential context can harm correctness despite a shorter reply.

Model price alone does not enforce output length or correctness.

Question 388

An application repeats a long, unchanged 5,000-token policy prefix in many calls to an eligible Bedrock model. It wants to reduce input-token cost through supported prompt caching. Which TWO implementation and verification steps are required?

Place the stable shared prefix before an appropriate model/API-supported cache checkpoint or enable supported automatic caching.
Assume cache hits are guaranteed merely because the same text appears, without checking usage fields.
Measure actual cached input tokens and cache writes in response usage against the same workload and billing terms.
Change the shared prefix every request to make the cache more personalized while expecting it to hit.
Use Bedrock batch inference as the caching mechanism for a multi-turn live conversation.

Explanation

Correct Answer

Caching reuses eligible stable prefixes; model, API, token minimum and cache-boundary support must be verified.

Usage fields distinguish reads and writes; net savings depend on real hit rates and model-specific cache pricing.

Incorrect Answers

Even eligible prefixes can miss because of TTL, size or changed content. Support does not guarantee a hit.

Changing the prefix changes its cache key, causing misses rather than reusing the same cached content.

Batch inference is asynchronous and not a substitute for supported interactive prompt caching.

Question 389

A prompt contains changing user input before static instructions. Which optimization balances actual cost and accepted task quality?

Place the changing username before the long static instruction block on every request.
Assume any two prompts with the same words in different order produce the same cached prefix.
Reorder static reusable prefix ahead of dynamic content to improve eligible cache hits.
Set cache TTL to an unsupported 1-hour value for every model without reading its card.

Explanation

Correct Answer

A stable static prompt prefix can be cached before varying user content; eligibility and actual reuse depend on the chosen model/API/cache settings.

Incorrect Answers

A varying prefix prevents stable checkpoint reuse after that position.

Cache keys depend on supported contiguous prompt content, not bag-of-words similarity.

TTL support is model-specific and may cause validation failure.

Question 390

A team assumes prompt caching always hits. Which resource-efficiency decision fits the workload?

Inspect actual cache usage metrics and model/API limits, not prompt repetition alone.
Equate two semantically similar prompts with a guaranteed explicit prompt-cache hit.
Assume a cached hit exists whenever the application logs a prompt repetition.
Use generation throughput improvement alone as a cache hit metric.

Explanation

Correct Answer

Measure cache-write/cache-read token counters and actual hit rate with supported model/API limits before estimating cost or latency savings.

Incorrect Answers

Checkpoint caching requires matching eligible token prefixes.

Model, minimum tokens and TTL can cause misses.

Faster output can have other causes and does not prove cache token reuse.

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

Question 380

Explanation

Correct Answer

Incorrect Answers

Question 381

Explanation

Correct Answer

Incorrect Answers

Question 382

Explanation

Correct Answer

Incorrect Answers

Question 383

Explanation

Correct Answer

Incorrect Answers

Question 384

Explanation

Correct Answer

Incorrect Answers

Question 385

Explanation

Correct Answer

Incorrect Answers

Question 386

Explanation

Correct Answer

Incorrect Answers

Question 387

Explanation

Correct Answer

Incorrect Answers

Question 388

Explanation

Correct Answer

Incorrect Answers

Question 389

Explanation

Correct Answer

Incorrect Answers

Question 390

Explanation

Correct Answer

Incorrect Answers

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