AWS AI Practitioner (AIF-C01) Practice Questions (2026): 10 Free Questions with Answers

Labelled examples (fraud / not fraud) used to predict a known target is the definition of supervised learning — here, binary classification. Unsupervised learning and clustering (A, D) have no labels and instead find hidden structure. Reinforcement learning (C) learns from reward signals through trial and error, not from a labelled dataset.

High training accuracy with poor performance on unseen data is classic overfitting — the model memorised noise instead of learning to generalise. Sensible fixes are more or more varied data, regularization, dropout, or a simpler model. Underfitting (A, D) shows low accuracy on both training and test data. Data drift (C) is a production issue where new data diverges over time, and raising the learning rate does not address generalization.

A foundation model is trained on massive, broad datasets and can be adapted to many tasks through prompting, retrieval-augmented generation, or fine-tuning. Option A describes a traditional task-specific model. Option C is a symbolic rules engine, not machine learning at all. Option D is unrelated infrastructure.

Embeddings map text into a vector space where semantically similar items sit close together, so a similarity search can match meaning rather than exact words. One-hot encoding (A) is sparse and captures no semantic relationship. Base64 (C) is just an encoding with no meaning. A primary key (D) is an identifier, not a semantic representation.

Amazon Bedrock is a fully managed, serverless service that offers foundation models from multiple providers — Anthropic, Amazon (Titan and Nova), Meta, Cohere, and others — behind one API. EC2 GPU instances (A) mean you manage the infrastructure yourself. SageMaker training jobs (C) are for building and training custom models. Amazon Comprehend (D) is a prebuilt NLP service, not a multi-provider foundation-model platform.

Lowering temperature reduces randomness, pushing the model toward the highest-probability tokens and more consistent, repeatable output. Raising temperature (A) increases variety and creativity — the opposite of what is needed. The max token limit (C) controls length, not determinism. Stop sequences (D) control where generation ends, not the consistency of the wording.

Retrieval-Augmented Generation retrieves relevant passages from your up-to-date documents at query time and supplies them to the model as context, so answers stay grounded and current — with citations — and no retraining is needed when documents change. Broad fine-tuning (A) is costly, quickly stale, and does not cleanly add private, changing data. Higher temperature (C) increases hallucination. Telling the model to “be accurate” (D) gives it no actual source facts.

Fine-tuning adjusts the model on curated examples so it durably changes default behaviour such as tone and style; both Amazon Bedrock and SageMaker JumpStart support model customization. RAG (B) supplies facts as context but does not reshape style. The token limit (C) affects length only. Output format (D) is unrelated to writing style.

Amazon Bedrock Guardrails let you define denied topics, content filters, and word and PII filters to block harmful or restricted output from a generative AI application. SageMaker Clarify (B) detects bias and helps explain model predictions, but it is not the runtime content filter for GenAI apps. Amazon Macie (C) discovers sensitive data in Amazon S3. CloudTrail (D) logs API activity for auditing.

AWS KMS handles encryption of data at rest (and TLS protects data in transit), while IAM identity-based policies enforce least-privilege access so only chosen roles can invoke specific models or actions. Public buckets and shared root credentials (A) break basic security hygiene. Disabling CloudTrail (C) removes the audit trail governance depends on. Hard-coding keys in source (D) is a serious, avoidable security risk.