30 Free AWS AI Practitioner Practice Questions with Answer Explanations

How These AWS AI Practitioner Practice Questions Are Organized

These 30 questions cover all five AIF-C01 exam domains: Fundamentals of AI and ML (Q1–Q6), Fundamentals of Generative AI (Q7–Q12), Applications of Foundation Models (Q13–Q18), Guidelines for Responsible AI (Q19–Q24), and Security, Compliance, and Governance (Q25–Q30). Each question includes the correct answer and a detailed explanation. The real exam has 65 questions (50 scored + 15 unscored), a 90-minute time limit, and a passing score of 700 out of 1,000. Registration is through aws.amazon.com/certification/certified-ai-practitioner for $100.

Domain 1: Fundamentals of AI and ML (Questions 1–6)

Q1. A retail company wants to segment its customers into groups based on purchasing behavior, without any predefined categories. Which type of machine learning is most appropriate?

Answer: Unsupervised learning (clustering). Unsupervised learning finds patterns in unlabeled data without predefined output categories. Supervised learning requires labeled training examples. The key phrase "without predefined categories" indicates unsupervised clustering.

Q2. An ML model achieves 98% accuracy on training data but only 71% accuracy on the validation dataset. What problem does this describe?

Answer: Overfitting. Overfitting occurs when a model learns the training data too specifically—including its noise—and fails to generalize to new data. Solutions include regularization, reducing model complexity, or collecting more diverse training data. The large gap between training and validation performance is the diagnostic signal.

Q3. A medical imaging company needs to detect rare diseases in X-ray scans. Missing a disease (false negative) is far more costly than a false alarm (false positive). Which evaluation metric should they prioritize?

Answer: Recall (also called sensitivity or true positive rate). Recall measures what proportion of actual positive cases the model correctly identifies. When false negatives are the most costly error type—as in disease detection—maximizing recall is the priority, even if it means accepting more false positives.

Q4. A company trains an ML model on customer service transcripts from the past five years, then deploys it to classify new tickets. The model performs significantly worse six months after deployment. What is the most likely cause?

Answer: Data drift (also called concept drift or distributional shift). The statistical properties of the input data have changed over time—new product categories, changed customer language, different ticket types—making the training data no longer representative of current inputs. Monitoring for data drift and retraining periodically addresses this.

Q5. Which of the following best describes the difference between artificial intelligence and machine learning?

Answer: Machine learning is a subset of AI in which systems learn patterns from data rather than following explicitly programmed rules. AI is the broader category of systems performing tasks that typically require human intelligence. ML is one specific approach within AI—learning from data. Other AI approaches include rule-based expert systems and search algorithms.

Q6. During the ML pipeline, a data scientist discovers that the training dataset contains 95% non-fraudulent transactions and only 5% fraudulent ones. What problem does this describe, and what is a common solution?

Answer: Class imbalance. Common solutions include oversampling the minority class (e.g., SMOTE), undersampling the majority class, or using class-weighted loss functions during training. A model trained on severely imbalanced data will learn to predict the majority class almost always, achieving high accuracy while failing at the actual task (detecting fraud).

Domain 2: Fundamentals of Generative AI (Questions 7–12)

Q7. A developer is building a customer support chatbot and wants it to always respond in a formal, professional tone. Which component of the prompt structure is most appropriate for setting this behavior?

Answer: A system prompt. System prompts set the model's overall behavior, persona, and constraints before any user interaction occurs. They persist across the conversation and are the correct mechanism for enforcing consistent tone, format, or behavioral guidelines.

Q8. A company's foundation model frequently produces outdated information about its product catalog, which changes weekly. The company does not want to retrain the model every week. Which approach addresses this?

Answer: Retrieval-augmented generation (RAG). RAG retrieves relevant, up-to-date documents at inference time and includes them in the prompt context, grounding the model's response in current information without modifying model weights. Fine-tuning would require weekly retraining—expensive and slow. RAG provides freshness without retraining.

Q9. Which scenario is best suited for fine-tuning a foundation model rather than using RAG?

Answer: A company wants the model to consistently use its internal legal terminology and response style across all outputs. Fine-tuning adapts model behavior and style by updating weights on domain-specific training data. It is best when the goal is changing how the model writes or reasons—not just what information it accesses. RAG is better for adding factual knowledge; fine-tuning is better for adapting behavior and tone.

Q10. A prompt reads: "Classify the following customer review as positive, neutral, or negative. Here are three examples: [example 1 - positive], [example 2 - neutral], [example 3 - negative]. Now classify: [review]." What prompting technique is this?

Answer: Few-shot prompting. Few-shot prompting provides multiple examples (two or more) of the desired task before the actual request. Three examples makes this few-shot (not zero-shot or one-shot). The examples guide the model on format and classification criteria.

Q11. A generative AI model produces a highly detailed but completely fabricated product specification for a product that does not exist. This information is presented confidently. What term describes this behavior?

Answer: Hallucination. LLMs generate statistically plausible continuations of text, which can produce confident-sounding false information. Hallucination is a known limitation of generative AI models. Mitigation approaches include RAG (grounding responses in verified documents) and human review before publication.

Q12. A model's context window is 8,192 tokens. A user submits a 10,000-token document for summarization. What happens?

Answer: The model cannot process the entire document in a single inference call—the input exceeds the context window limit. Solutions include chunking the document into segments small enough for the context window, using a model with a larger context window, or using a map-reduce summarization approach that summarizes chunks and then combines the summaries.

Domain 3: Applications of Foundation Models (Questions 13–18)

Q13. A company needs to extract structured data—invoice numbers, dates, and line items—from thousands of scanned PDF invoices. Which AWS service is most appropriate?

Answer: Amazon Textract. Textract is purpose-built for extracting text and structured data from scanned documents, forms, and PDFs—including tables and key-value pairs. Rekognition handles image analysis but is not optimized for document data extraction. Comprehend extracts entities from text but requires text input, not raw document images.

Q14. A company wants to build a chatbot that answers questions about its internal knowledge base using natural language. The knowledge base is stored as thousands of Word documents. Which combination of AWS services is most appropriate?

Answer: Amazon Kendra (for document indexing and retrieval) + Amazon Bedrock (for response generation). This is a RAG architecture: Kendra retrieves relevant document passages in response to a query; Bedrock generates a natural language answer grounded in those passages. Amazon Lex handles conversation flow if needed.

Q15. A media company needs to automatically moderate user-uploaded images to detect explicit content at scale. Which AWS service handles this?

Answer: Amazon Rekognition. Rekognition includes a content moderation feature that detects explicit or suggestive content in images and videos, with confidence scores. It is purpose-built for this use case and requires no custom ML development.

Q16. A company is using Amazon Bedrock and wants to give the foundation model access to a company-specific knowledge base for RAG. Which Bedrock feature enables this?

Answer: Amazon Bedrock Knowledge Bases. Bedrock Knowledge Bases is the managed RAG feature within Amazon Bedrock. It connects a vector store (containing embedded document chunks) to a foundation model, enabling grounded responses without custom retrieval infrastructure.

Q17. A developer wants to use a foundation model through Amazon Bedrock but needs the model to take multi-step actions—searching the web, calling an API, and writing a summary—autonomously. Which Bedrock feature supports this?

Answer: Amazon Bedrock Agents. Bedrock Agents enables orchestrating multi-step tasks using foundation models. The agent can reason about what action to take next, call defined action groups (APIs, Lambda functions), and synthesize a final response—without the developer writing explicit orchestration logic for each step.

Q18. Which AWS service would you use to convert spoken customer service calls to text for sentiment analysis at scale?

Answer: Amazon Transcribe. Transcribe converts speech to text, including support for call analytics (identifying sentiment, non-talk time, and interruptions in call recordings). The resulting text can then be processed by Amazon Comprehend for detailed sentiment analysis.

Domain 4: Guidelines for Responsible AI (Questions 19–24)

Q19. An AI hiring tool consistently rates female candidates lower for engineering roles than male candidates with identical qualifications. The model was trained on ten years of historical hiring decisions. What type of bias is this?

Answer: Historical bias (also called data bias). The model learned patterns from historical decisions that themselves reflected gender bias in hiring. The model replicated and encoded this historical bias. The root cause is not the algorithm—it is the training data reflecting biased past decisions.

Q20. A bank uses an ML model for loan approval. A rejected applicant requests to know why they were denied. Which AWS tool provides the explanation of which features most influenced the model's decision?

Answer: Amazon SageMaker Clarify. SageMaker Clarify generates SHAP (SHapley Additive exPlanations) values that quantify each feature's contribution to a specific prediction. This supports explainability requirements—the ability to tell a rejected applicant that "income and credit history were the two most influential factors in your denial."

Q21. A company deploys an AI model for medical triage that flags high-risk patients for immediate physician review. A patient is flagged incorrectly as low-risk and receives delayed care. Which responsible AI principle does improving the oversight of this system address?

Answer: Human oversight (also called controllability). High-stakes decisions in healthcare require human review of model outputs before action is taken. Amazon Augmented AI (A2I) provides the infrastructure for human review workflows in AI pipelines.

Q22. Which of the following is the best description of "fairness" as a dimension of AWS's responsible AI principles?

Answer: Ensuring AI systems produce equitable outcomes across different demographic groups and do not encode or amplify discriminatory patterns. Fairness in AI means evaluating model performance across subgroups—not just overall accuracy—to identify and mitigate differential impacts on protected groups.

Q23. A company's AI model makes decisions that are difficult to explain to affected users. Which AWS feature helps identify why the model produced a specific output?

Answer: Amazon SageMaker Clarify (for feature importance and bias detection) and SageMaker Model Cards (for documenting model behavior and intended use). Clarify provides quantitative feature importance; Model Cards provide qualitative documentation of model purpose, performance, and limitations—both support explainability obligations.

Q24. When should a company use Amazon Augmented AI (A2I) in an AI workflow?

Answer: When model predictions need human review before a consequential decision is made—particularly for low-confidence predictions or high-stakes use cases. A2I routes model outputs to human reviewers when the model's confidence falls below a threshold or for a random sample of outputs for quality monitoring. It integrates with Textract and Rekognition natively.

Domain 5: Security, Compliance, and Governance (Questions 25–30)

Q25. A company stores sensitive training data in Amazon S3 and wants to detect if any personally identifiable information (PII) is present in those datasets. Which AWS service handles this automatically?

Answer: Amazon Macie. Macie uses ML to automatically discover and classify sensitive data in S3, including PII (names, SSNs, credit card numbers). It produces findings that help data teams remediate privacy risks before using the data for model training.

Q26. A company needs a complete audit trail of all API calls made to Amazon Bedrock for compliance purposes. Which AWS service provides this?

Answer: AWS CloudTrail. CloudTrail logs all API calls made to AWS services, including Bedrock, with timestamps, caller identity, and parameters. This audit log is required for compliance frameworks that mandate accountability for AI system usage.

Q27. A data science team wants to manage versions of their trained ML models, require approval before promoting a model to production, and track model metadata and lineage. Which AWS service supports all of this?

Answer: Amazon SageMaker Model Registry. Model Registry provides a centralized catalog for ML models with versioning, approval workflows (pending/approved/rejected), and metadata tracking. It is the model governance layer in AWS ML pipelines.

Q28. An organization wants to ensure its AWS environment remains compliant with its security policies as AI resources are provisioned and changed. Which service continuously monitors resource configurations for compliance?

Answer: AWS Config. AWS Config continuously records resource configurations and evaluates them against rules that encode desired compliance states. It can alert when a resource drifts out of compliance—for example, if an S3 bucket containing training data is made public.

Q29. A healthcare company is deploying an AI model that processes patient data. Which compliance framework governs the handling of patient health information in the United States?

Answer: HIPAA (Health Insurance Portability and Accountability Act). HIPAA governs protected health information (PHI) in the US. AWS provides a HIPAA Eligible Services list and a Business Associate Agreement (BAA) for customers processing PHI. AI workloads involving patient data must use HIPAA-eligible services and implement required security controls.

Q30. A company wants to control which IAM roles have permission to invoke specific Amazon Bedrock foundation models, and to log all invocations for audit purposes. Which two AWS tools address these requirements?

Answer: AWS IAM (for permission control via resource-based and identity-based policies) and AWS CloudTrail (for logging all invocation API calls). IAM policies restrict who can call Bedrock's InvokeModel API for specific model ARNs. CloudTrail captures every invocation with caller identity, timestamp, and model used.

Exam details verified against aws.amazon.com/certification/certified-ai-practitioner as of 2026-02-27. Fees and requirements are subject to change — confirm current details at aws.amazon.com/certification/certified-ai-practitioner before your exam date.