AWS AI Practitioner (AIF-C01) — Practice Exam

Domain 1 · Question 1

Which statement best describes machine learning relative to artificial intelligence?

A. ML is unrelated to AI; it only applies to databases. B. ML is a subset of AI where behavior improves from data without relying only on hand-written rules. C. AI is a subset of ML focused on neural networks only. D. Deep learning is the same as classical rules-based expert systems.

Domain 1 · Question 2

A model predicts housing prices from labeled square footage and zip features. This is best described as:

A. Unsupervised clustering B. Supervised regression C. Reinforcement learning D. Batch inference only

Domain 1 · Question 3

Which pair correctly contrasts batch and real-time inference?

A. Batch optimizes per-request latency; real-time optimizes total records per night. B. Batch scores many records on a schedule; real-time serves low-latency API requests. C. They are identical terms in AWS documentation. D. Real-time never uses GPUs.

Domain 1 · Question 4

In the ML lifecycle, which stage typically happens immediately after model evaluation in many pipelines?

A. Data collection B. Hyperparameter tuning C. Deployment D. Exploratory data analysis

Domain 1 · Question 5

MLOps primarily aims to:

A. Replace DevOps entirely for all companies. B. Make ML reproducible, monitorable, and production-ready with pipelines and governance. C. Guarantee 100% model accuracy. D. Remove the need for test data.

Domain 1 · Question 6

F1 score is most useful when you need to balance:

A. Training time and inference cost B. Precision and recall on a classification task C. S3 bucket size and Region count D. CPU vs GPU clock speed

Domain 1 · Question 7

Which AWS service is a managed NLP API for text analysis tasks such as entities and sentiment?

A. Amazon Polly B. Amazon Comprehend C. Amazon Textract D. AWS CloudTrail

Domain 1 · Question 8

Concept drift means:

A. The training code failed to compile. B. The relationship between inputs and targets changes over time, so the old model becomes misaligned. C. The VPC subnet ran out of IP addresses. D. Temperature was set too low on the FM.

Domain 1 · Question 9

Which task is OUT of scope for the AWS Certified AI Practitioner foundational exam objective for implementers?

A. Recognizing when ML is appropriate for a business case B. Describing the ML lifecycle at a high level C. Building and deploying a full custom training pipeline as the primary hands-on expectation D. Understanding batch vs real-time inference

Domain 1 · Question 10

Amazon SageMaker is best described as:

A. A speech synthesizer only B. A broad AWS platform for building, training, deploying, and monitoring custom ML C. A billing-only dashboard D. A DNS service

Domain 2 · Question 11

A foundation model (FM) is best described as:

A. A fixed rules engine for tax calculation B. A broad base model often pre-trained, adaptable via prompting, RAG, fine-tuning, or agents C. A single Excel macro D. Only usable for image classification with no text

Domain 2 · Question 12

Why is chunking commonly used before embedding long documents?

A. Chunking removes the need for vectors B. Chunks fit context limits and improve retrieval granularity for RAG-style patterns C. Chunking is only for video files D. AWS forbids embedding without chunking

Domain 2 · Question 13

Amazon Bedrock is primarily:

A. A relational database engine B. A managed service for accessing and customizing foundation models and related GenAI features C. An IoT protocol D. A mainframe emulator

Domain 2 · Question 14

Higher temperature in FM sampling typically:

A. Forces identical outputs every run B. Increases randomness and output diversity (and can increase hallucination risk) C. Disables token billing D. Only affects S3 storage class

Domain 2 · Question 15

Token-based pricing for GenAI APIs typically charges by:

A. Only the number of S3 buckets B. Input and output tokens consumed by the model C. Number of Route 53 hosted zones D. RDS Multi-AZ boolean flag

Domain 2 · Question 16

Which option is a hallmark limitation of generative AI for factual enterprise use cases?

A. Hallucinations and unsupported claims without grounding B. It cannot generate text C. It only runs on-premises with no cloud option D. It never uses GPUs

Domain 2 · Question 17

SageMaker JumpStart is best described as:

A. A curated way to start with models, solutions, and notebooks inside SageMaker B. A VPC flow log parser only C. A hardware store for GPUs D. A replacement for IAM

Domain 2 · Question 18

Which modality pair is most associated with a multimodal foundation model?

A. Only CSV and JSON schema B. Text and images (or other cross-modal inputs/outputs) C. ICMP and UDP D. CloudFormation templates only

Domain 2 · Question 19

Provisioned throughput for a managed FM is most about:

A. Deleting all logs B. Reserving capacity for predictable latency and spend versus pure on-demand bursts C. Disabling encryption D. Forcing batch-only inference

Domain 2 · Question 20

RLHF is most closely associated with:

A. Routing Lambda functions across Regions B. Aligning model behavior to human preference signals as part of training/fine-tuning C. Backing up EBS volumes D. Compiling C++ kernels

Domain 2 · Question 21

Amazon Q (in AWS positioning) is best described as:

A. A DNS resolver B. An assistant-style product for builder/business productivity integrated with AWS and enterprise tools C. A graph database D. A Snowball appliance

Domain 2 · Question 22

Business metrics for GenAI pilots should often include items such as:

A. Only BLEU score B. ROI, handle time, conversion, or cost per resolution—not only technical scores C. Only VPC subnet IDs D. Only EC2 instance families

Domain 3 · Question 23

Retrieval-Augmented Generation (RAG) primarily improves factual grounding by:

A. Removing the foundation model from the pipeline B. Retrieving relevant chunks and conditioning generation on that context C. Disabling embeddings D. Running only batch jobs

Domain 3 · Question 24

Which AWS exam-guide example category can host vector search for embeddings?

A. Only AWS IoT Core rules B. Amazon OpenSearch Service and managed database options that support vector patterns (e.g. RDS/Aurora Postgres with pgvector) C. Amazon Chime SDK only D. AWS Certificate Manager only

Domain 3 · Question 25

Agents for Amazon Bedrock most directly support:

A. Static S3 website hosting without APIs B. Multi-step workflows that plan, call tools/APIs, and use retrieval C. Replacing KMS entirely D. Legacy COBOL transpilation only

Domain 3 · Question 26

Increasing max output tokens primarily affects:

A. Only S3 object tags B. The upper bound on generated length (and often cost) C. VPC route table propagation speed D. KMS key rotation period

Domain 3 · Question 27

Few-shot prompting means:

A. Providing zero examples in the prompt B. Supplying a small number of labeled examples inside the prompt to steer formatting/behavior C. Running zero training epochs D. Deleting the knowledge base

Domain 3 · Question 28

Prompt injection risk is best mitigated by a combination of controls such as:

A. Only using larger models with no other changes B. Isolation, least-privilege tool access, guardrails, and secure prompt/system boundaries C. Publishing root access keys in the prompt D. Disabling CloudTrail

Domain 3 · Question 29

Instruction tuning typically refers to:

A. Tuning CPU governors on EC2 B. Supervised fine-tuning on instruction–response pairs to improve helpfulness/format adherence C. Only changing S3 lifecycle policies D. Patching the hypervisor

Domain 3 · Question 30

ROUGE is most associated with evaluating:

A. Summarization/text overlap-oriented quality B. RDS storage autoscaling C. Direct Connect redundancy D. WAF rule counts

Domain 3 · Question 31

BERTScore is most associated with:

A. Semantic similarity using contextual embeddings versus strict n-gram overlap B. Subnet sizing C. EBS throughput limits D. Route 53 health check intervals only

Domain 3 · Question 32

Cost tradeoff: fine-tuning a custom FM versus RAG-only often implies:

A. RAG always has zero engineering cost B. Fine-tuning can improve specialization but raises MLOps burden and hosting complexity versus retrieval-first approaches C. Fine-tuning removes the need for evaluation D. RAG forbids private VPCs

Domain 3 · Question 33

Top-p (nucleus) sampling is a technique to:

A. Encrypt prompts at rest B. Truncate sampling mass to likely tokens to control diversity C. Provision NAT gateways D. Create IAM roles automatically

Domain 3 · Question 34

When selecting an FM, context length is important because:

A. It sets the maximum prompt + completion window available in one call B. It defines S3 maximum object size C. It replaces GDPR compliance D. It is always equal to 256 bytes

Domain 3 · Question 35

Jailbreaking in GenAI typically refers to:

A. Using multi-factor authentication B. Attempts to bypass safety policies or elicit disallowed outputs C. Migrating data to Glacier D. Enabling VPC Flow Logs

Domain 3 · Question 36

Human evaluation for FM quality is valuable especially for:

A. Situations where automated metrics do not capture helpfulness, safety, or business fit B. Replacing all logging C. Avoiding any documentation D. Eliminating IAM

Domain 4 · Question 37

Amazon SageMaker Clarify is primarily used for:

A. Billing reports only B. Bias and explainability analysis for models and data (exam-relevant capability) C. DNS failover D. Certificate issuance

Domain 4 · Question 38

Guardrails for Amazon Bedrock is best described as:

A. A tool to remove all encryption B. Policy controls such as content filters and PII/topic constraints for GenAI applications C. A network load balancer D. A Hadoop cluster

Domain 4 · Question 39

Subgroup analysis in responsible AI helps:

A. Guarantee equal salaries for all employees automatically B. Surface disparate performance across segments that a global metric may hide C. Disable model monitoring D. Replace the need for datasets

Domain 4 · Question 40

A Model Card commonly documents:

A. Only the model's favorite color B. Intended use, data, evaluation metrics, limitations, and ethical considerations C. Only EC2 pricing D. IAM password rotation schedules only

Domain 4 · Question 41

Amazon Augmented AI (A2I) supports:

A. Human review workflows for predictions that need oversight B. Physical warehouse robots only C. Replacing SageMaker D. Deleting CloudTrail logs

Domain 4 · Question 42

A key tension between stricter safety filtering and interpretability can be:

A. Stronger safety stacks may reduce transparency into why outputs were blocked or altered B. There is never any tradeoff C. Safety filters always increase interpretability D. Interpretability removes safety needs

Domain 4 · Question 43

Veracity in responsible AI most closely relates to:

A. Subnet CIDR design B. Truthfulness of outputs relative to evidence and claims C. S3 request rate D. Lambda concurrency caps only

Domain 5 · Question 44

Under the shared responsibility model, which is generally a customer responsibility for AI apps on AWS?

A. Physically securing data centers worldwide B. Configuring IAM least privilege for roles that invoke Bedrock and access training data C. Patching the hypervisor only by AWS field technicians with no customer input D. Operating regional power grids

Domain 5 · Question 45

Amazon Macie is most useful for:

A. Discovering and protecting sensitive data such as PII in S3-scale repositories B. Issuing TLS certificates C. Compiling deep learning kernels D. Managing ECS task placement

Domain 5 · Question 46

AWS PrivateLink helps secure connectivity by:

A. Providing private access to many AWS services without traversing the public internet B. Deleting all IAM policies C. Forcing plaintext HTTP D. Disabling VPC endpoints entirely

Domain 5 · Question 47

AWS CloudTrail is primarily for:

A. API audit logging and governance of account activity B. Generating embeddings C. Training transformers from scratch D. Purchasing Savings Plans only

Domain 5 · Question 48

AWS Audit Manager is best described as:

A. Helping collect continuous audit evidence aligned to controls and frameworks B. A relational migration tool only C. A game engine D. An image generator

Domain 5 · Question 49

Encrypting model artifacts at rest with KMS is an example of:

A. Customer responsibility for data protection in many architectures B. AWS-only responsibility with no customer keys possible C. A alternative to using any IAM roles D. A replacement for network segmentation

Domain 5 · Question 50

Data residency policies in governance most directly concern:

A. Which geographic Region(s) data may be stored and processed in B. The color of the Bedrock console C. Lambda memory size D. Number of Git branches

Practice exam · 50 questions

Answer key