AIF-C01 mock exam: 25 free practice questions

Q1 / 25

A banking assistant receives user text that says to ignore its developer instructions and act as a chef. Which risk best describes the attempt?

1. AHijacking through prompt injection
2. BExposure through prompt leakage
3. CPoisoning of the training dataset
4. DJailbreaking of native safety controls

Q2 / 25

At what stage does context engineering primarily change what a model receives?

1. ADuring model pretraining
2. BWhile retrieval searches for external material
3. CAt inference time for a request
4. DWhile memory stores earlier interaction state

Q3 / 25

A team must score hundreds of generated summaries against approved reference summaries. The result must be deterministic, and the team specifically wants lexical coverage rather than contextual similarity. Which metric best fits?

1. AA second language model that scores and explains each summary
2. BBERTScore applied to contextual embeddings of both summaries
3. CHuman review focused on coherence and user experience
4. DROUGE applied to generated and reference summaries

Q4 / 25

What is the role of SageMaker JumpStart in generative AI development?

1. AA model hub for evaluating, selecting, customizing, and integrating models
2. BA managed runtime for isolating and scaling agent sessions
3. CA conversational suite for analytics, workflows, apps, and research
4. DAn agentic IDE that turns specifications into implementation plans

Q5 / 25

Which metric measures the share of all classification predictions that are correct?

1. AAccuracy across the complete prediction set
2. BPrecision among the predicted positive cases
3. CCost per user across the served population
4. DRecall among the actual positive cases

Q6 / 25

Which prompt component tells the model what action to perform?

1. AThe negative prompt
2. BThe instruction
3. CTask context
4. DThe demonstration count

Q7 / 25

Under a token-based pricing option, how is inference cost calculated?

1. AThe displayed paragraph count
2. BThe number of model providers listed in the service catalog
3. CApply the relevant input and output token rates to the respective token quantities for the selected model and inference option.
4. DThe number of evaluation criteria written before deployment

Q8 / 25

Which statement correctly compares data properties?

1. ALabeling describes whether targets are supplied, while structure describes whether data follows a fixed schema
2. BLabeling describes rows and columns, while structure describes whether targets are correct
3. CLabeling describes chronological order, while structure describes whether an image contains a caption
4. DLabeling describes text modality, while structure describes whether observations have timestamps

Q9 / 25

Which cost placement best distinguishes pre-training from request-time guidance?

1. APre-training adds examples to each request without training
2. BPre-training incurs model-training work before deployment
3. CPre-training retrieves private chunks only during inference
4. DPre-training creates no upfront work and only caches prompts

Q10 / 25

Why are hallucinations a distinct risk from biased outputs?

1. AHallucinations concern protected works; bias concerns excessive computational resource use.
2. BHallucinations concern customer relationships; bias concerns human approval before release.
3. CHallucinations concern model reuse; bias concerns data storage and processing efficiency.
4. DHallucinations concern fabricated or incorrect information; bias concerns disparities across groups.

Q11 / 25

Which Amazon Quick feature connects an organization's documents and data sources to supply the information that grounds answers?

1. AAmazon Quick Index
2. BAmazon Quick Flows
3. CAmazon Quick Sight
4. DAmazon Quick Research

Q12 / 25

A model scores each case and an investigator then decides what to do with it. What is AI/ML contributing?

1. AA replacement for the investigator's accountability
2. BA written transcript of the investigator's own notes
3. CAssistance for a human decision, supplied as evidence
4. DAn exact result already prescribed by the applicable rules

Q13 / 25

What is the expected response effect of increasing temperature?

1. ALower-probability token choices become more likely
2. BThe maximum output-token count is reduced
3. CThe submitted context becomes shorter
4. DGeneration stops at a specified character sequence

Q14 / 25

Users understand the capability guide before they start, but screen-reader users cannot submit their inputs, the interface stays silent while the request is processed, and finished results cannot be contested. Which release package repairs the journey?

1. AScreen-reader-accessible input controls, an estimated duration shown before submission, and result explanations with a challenge channel.
2. BScreen-reader-accessible input controls, progress information during processing, and result explanations with a challenge channel.
3. CScreen-reader-accessible input controls, progress information during processing, and result explanations with confidence scores.
4. DA more detailed capability guide, progress information during processing, and result explanations with a challenge channel.

Q15 / 25

A team runs an AI application on Amazon EC2 and installs its own inference software. It must assign responsibility for operating-system patches and application updates. Who owns both tasks?

1. AThe customer, because both layers run inside its EC2 workload
2. BAWS, because EC2 transfers guest software maintenance to the provider
3. CThe external auditor, because patching is a compliance control
4. DAWS for the application and the customer for physical infrastructure

Q16 / 25

Why can adding unrelated material weaken a request context?

1. AIt permanently replaces the model's learned parameters
2. BIt makes retrieval unnecessary for every application task
3. CIt converts conversation state into learned parameters
4. DIt can compete with the task-relevant information

Q17 / 25

A contact center wants an agent-assist panel that shows, while the call is still running, what the caller is saying and whether the caller sounds negative. Which combination delivers that?

1. AComprehend on the live stream, then Transcribe on the text
2. BLex on the live stream, then Polly on the sentiment result
3. CTranscribe on the live stream, then Comprehend on the text
4. DTranslate on the live stream, then Comprehend on the text

Q18 / 25

A chat assistant answers well, but its serving cost and latency at peak traffic are unacceptable, and leadership wants the replacement to behave as closely as it can to the current model. The team accepts one extra training stage. Which approach fits?

1. AContinue pre-training the current model with newly collected traffic
2. BDistill the current model into a smaller student trained on its responses
3. CInstruction tune the current model on labeled prompt-and-output pairs
4. DServe a smaller untuned model and offset the quality gap with longer few-shot prompts

Q19 / 25

What does AI decision transparency provide to a user?

1. AInformation about system activity, confidence, and the basis for an output
2. BA channel that collects challenges without explaining the decision
3. CA performance score with no context about model behavior
4. DA guarantee that every model output is factually correct and harmless for every request

Q20 / 25

Which IAM feature gives an AWS workload temporary credentials without distributing long-term IAM user credentials?

1. AAn IAM role assumed by the workload
2. BA Macie sensitive-data discovery job
3. CA PrivateLink endpoint policy used as a password
4. DAn IAM policy stored as the workload credential

Q21 / 25

Which generative AI advantage supports different tasks or contexts through supplied input and configuration?

1. AAdaptability
2. BInaccuracy
3. CNondeterminism
4. DInterpretability limitation

Q22 / 25

A bank wants to score each card payment for suspicion as it happens, and separately to plan next year's monthly transaction volume. Which pairing fits the two needs?

1. AClustering now; a knowledge base for the yearly plan
2. BForecasting for both, since both read past behavior
3. CFraud detection now; forecasting for the yearly plan
4. DFraud detection for both, since both score payments

Q23 / 25

A RAG assistant begins returning attacker-controlled claims after a newly added repository document is retrieved. The system prompt remains secret, and users are not overriding instructions. Which diagnosis matches both the affected asset and stage?

1. APrompt exposure that reveals developer instructions
2. BGoal hijacking that redirects the system instruction
3. CReference-data poisoning that influences grounded responses
4. DJailbreaking that bypasses native content moderation

Q24 / 25

A classifier performs poorly on training data and evaluation data. The team confirms that labels are reliable and that demographic subgroup metrics are comparable. It must diagnose model fit before changing production monitoring. Which diagnosis best fits?

1. AOverfitting because the model memorizes training data but cannot generalize
2. BUnderfitting because the model fails to capture the training relationship
3. CBias drift because a live metric exceeded its configured baseline threshold
4. DLabel corruption because actual outcomes cannot be compared with predictions

Q25 / 25

An organization must process model requests inside the EU and uses a Service Control Policy that permits only its source Region. Calls made through the EU geographic inference profile start to fail. What explains the failure and resolves it?

1. AThe data no longer resides in the source Region, so it has to be replicated to every destination Region before calls succeed
2. BEncryption in transit is blocking the routed requests, so it has to be disabled between the source and destination Regions
3. CThe profile can route to other destination Regions, so the policy must allow them or carry an inference-profile exception
4. DGeographic profiles send data outside the EU, so the workload has to stop using cross-Region inference to stay compliant