AWS CertifiedAIF-C01 Exam

Mock Exam AI Practitioner

Prepare for the AWS AI Practitioner (AIF-C01) exam with practical questions and detailed explanations.

Free

Detailed explanations

Updated 2026

This mock exam is an independent educational resource. It is not affiliated with, endorsed by, or sponsored by Amazon Web Services (AWS).

Practice by domain

How do you want to train?

Choose the mode that fits your current goal.

Study Mode

Feedback after every answer, free navigation, no timer.

✓Full explanation after each answer
✓Go back and review at any time
✓No time limit

Exam Mode

Realistic simulation of the actual exam with a countdown.

✓90 minutes to answer everything
✓No intermediate feedback
✓Score and answer key only at the end

Questions and commented answer key

See all questions with the correct answers and detailed explanations to reinforce your learning.

Questions

Topics

AI Definition

1 question

Question 1
What is Artificial Intelligence (AI) in its most basic definition?
Correct answer: B. Computers performing tasks that would normally require human intelligence
AI is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as pattern recognition, decision-making, and natural language processing. It is not just storage, specific hardware, or a programming language.

ML Types

1 question

Question 2
Which type of machine learning uses labeled data (with the correct answer) to train a model to make predictions?
Correct answer: A. Supervised learning
Supervised learning uses labeled data (input → expected output) to train classification or regression models. Unsupervised learning finds patterns in unlabeled data (clustering), reinforcement learning learns through trial and error with rewards, and deep learning is a technique (deep neural networks) applicable to any of these types.

ML Pipeline

1 question

Question 3
In a typical machine learning pipeline, what is the difference between the training and inference phases?
Correct answer: B. Training adjusts model parameters from data; inference uses the trained model to make predictions on new data
Training is the phase where the model learns by adjusting its parameters (weights) from historical data. Inference (also called prediction) is the phase where the trained model processes new data to generate results. The two phases may use different hardware, but the core distinction is that training creates/updates the model, while inference uses it.

AWS ML Services

1 question

Question 4
Which AWS service is a fully managed platform to develop, train, and deploy machine learning models at scale?
Correct answer: B. Amazon SageMaker
Amazon SageMaker is the AWS managed platform for the full ML lifecycle: development (Studio, notebooks), training, deploying models to endpoints, and monitoring. Lambda is serverless compute, Comprehend is a pre-trained NLP service for specific tasks, and Glue is data ETL.

GenAI Concepts

1 question

Question 5
What differentiates Generative AI from traditional AI?
Correct answer: B. Generative AI creates new content (text, image, code, audio) instead of just classifying or predicting
Generative AI creates new content (text, images, code, audio, video) based on patterns learned during training, while traditional AI focuses on classifying, predicting, or analyzing existing data. It uses ML models (especially foundation models), works with various data types, and can run on any platform.

Foundation Models

1 question

Question 6
What is a Foundation Model (FM) in the context of generative AI?
Correct answer: B. A large-scale model pre-trained on massive amounts of data that can be adapted for various downstream tasks
Foundation Models are AI models pre-trained on massive amounts of data (texts, images, etc.) that serve as a "foundation" to be adapted to multiple tasks via fine-tuning, RAG, or prompt engineering. Examples: Claude, Llama, Titan, Stable Diffusion. They are not instances, frameworks, or single-task models.

Tokens

1 question

Question 7
In the context of Large Language Models (LLMs), what is a token?
Correct answer: B. The basic unit of text the model processes, typically a word or part of a word
Tokens are the basic units LLMs process — they can be whole words, sub-words, or characters depending on the tokenizer. Usage cost and model context limits are measured in tokens. Not to be confused with authentication tokens or encryption keys.

Prompts

1 question

Question 8
What is a prompt in the context of generative AI?
Correct answer: B. The input (text, image, or other data) provided to the model to guide it in generating a response
A prompt is the input (usually text) the user provides to the model to guide its generation — it may include instructions, context, examples, and questions. Prompt quality directly affects response quality (hence the discipline of "prompt engineering"). It is not an instance, password, or model weight.

Amazon Bedrock

1 question

Question 9
Which AWS service provides managed access to foundation models from multiple providers (Anthropic, Meta, Cohere, Amazon, etc.) through a unified API?
Correct answer: B. Amazon Bedrock
Amazon Bedrock is the AWS managed service that provides access to foundation models from various providers via a single API, without having to manage infrastructure. SageMaker is a general ML platform, Lambda is serverless compute, and Comprehend is pre-trained NLP for specific tasks.

Prompt Engineering

2 questions

Question 10
Which prompt engineering technique involves providing the model with a few examples of (input, expected output) pairs before the actual question to guide the response style?
Correct answer: B. Few-shot prompting
Few-shot prompting provides the model with a few examples (typically 2-5) inside the prompt so it learns the expected pattern. Zero-shot is without examples (instruction only), fine-tuning retrains model weights with new data (not prompt engineering), and pre-training is the original foundation model creation phase.
Question 42
Which prompt engineering technique asks the model to "think step by step" before giving the final answer, improving performance on reasoning problems?
Correct answer: A. Chain-of-Thought (CoT) prompting
Chain-of-Thought (CoT) prompting instructs the model to explicitly reason step by step before the final answer. Particularly useful for math, logic, and planning problems. Zero-shot is instruction without examples, token streaming is incremental token delivery, and quantization is a model compression technique.

RAG

1 question

Question 11
Which technique allows a foundation model to answer using company-specific, up-to-date information without retraining the model?
Correct answer: B. Retrieval Augmented Generation (RAG)
RAG (Retrieval Augmented Generation) retrieves relevant documents from a knowledge base (typically a vector database such as Bedrock Knowledge Bases) and injects the content into the prompt before generation, letting the model answer using specific data without retraining. Pre-training creates the model from scratch, quantization reduces size, and tokenization splits text into tokens.

Fine-Tuning

1 question

Question 12
When does it make more sense to apply fine-tuning to a foundation model instead of using prompt engineering?
Correct answer: B. When you need to adapt the model to a very specific domain with its own vocabulary and have quality labeled data
Fine-tuning is appropriate when the use case requires deep adaptation to a specific domain (technical jargon, consistent format, very specific behavior) and quality training data exists. It is more expensive and time-consuming than prompt engineering, so it should only be used when prompt engineering is insufficient. For light tone or style adjustments, prompt engineering is enough.

Amazon Q

1 question

Question 13
Which AWS service is a generative AI assistant for work that can answer questions, generate content, and automate tasks, focused on enterprise environments?
Correct answer: A. Amazon Q
Amazon Q is the AWS generative AI assistant focused on enterprise use cases (Q Developer for code, Q Business for company knowledge). Lex is for building custom chatbots, Polly does text-to-speech, and Comprehend is pre-trained NLP for text analysis.

Bias

1 question

Question 14
What is "bias" in Artificial Intelligence systems?
Correct answer: B. Systematic tendencies that lead to unfair or inaccurate results for certain groups, typically inherited from training data
AI bias refers to systematic tendencies in the model that produce unequal or unfair results for specific groups (gender, race, age, etc.). It typically comes from training data that reflects historical inequalities or imbalanced sampling. Mitigating bias is a central pillar of responsible AI.

AI Service Cards

1 question

Question 15
What are AWS AI Service Cards?
Correct answer: B. Documents published by AWS that describe intended use cases, limitations, responsible design considerations, and best practices for each AI service
AWS AI Service Cards are transparency documents published by AWS for AI services (such as Rekognition Face Liveness, Bedrock FMs, etc.), describing intended use cases, limitations, responsible design considerations, and best practices. They serve as a guide for ethical service use.

Encryption

1 question

Question 16
How does Amazon Bedrock protect data that customers send to foundation models by default?
Correct answer: B. Data is encrypted in transit (TLS) and at rest, and is not used to train other customers’ models
Amazon Bedrock encrypts all data in transit (TLS) and at rest, keeps data isolated within the customer’s account, and does NOT use prompt or response content to train foundation models or share with third parties. This is a core data governance principle of Bedrock.

IAM

1 question

Question 17
Which AWS security practice should be used to grant an application access to Amazon SageMaker without exposing permanent credentials?
Correct answer: B. Use IAM Roles with temporary permissions
IAM Roles are the AWS recommended practice for granting access to services (including SageMaker) because they provide rotating temporary credentials and follow the principle of least privilege. Sharing root credentials, hardcoding, or disabling IAM directly violate AWS security best practices.

Deep Learning

1 question

Question 18
Which machine learning technique uses neural networks with multiple layers to learn complex data representations, serving as the foundation of many generative AI models?
Correct answer: C. Deep Learning (deep neural networks)
Deep Learning uses neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data. It is the foundation of modern AI models, including CNNs (vision), RNNs/Transformers (language), and foundation models. Linear regression, decision trees, and Naive Bayes are simpler ML techniques that do not involve multiple neural layers.

Classification vs Regression

1 question

Question 19
What is the difference between a classification model and a regression model in machine learning?
Correct answer: A. Classification predicts discrete categories (e.g., spam vs. not-spam); regression predicts continuous numeric values (e.g., price)
Classification predicts discrete categories (spam/not-spam, sick/healthy, cat/dog). Regression predicts continuous numeric values (price, temperature, age). The choice depends on the target variable type. It has nothing to do with speed or input type — both can use images, numbers, or text as inputs.

Model Evaluation

1 question

Question 20
In a binary classification problem with highly imbalanced data (e.g., fraud detection), which metric is generally more informative than simple accuracy?
Correct answer: B. F1-score (combination of precision and recall)
F1-score combines precision (correct predictions among predicted positives) and recall (actual positives detected), being more informative than accuracy on imbalanced datasets. In fraud detection, a model that always predicts "non-fraud" might achieve 99% accuracy but be useless. The other options (time, size, epochs) are operational details, not model quality metrics.

Embeddings

1 question

Question 21
What are embeddings (vector representations) in the context of generative AI?
Correct answer: B. Numeric vectors that represent the semantic meaning of texts, images, or other data, enabling similarity comparisons
Embeddings are high-dimensional numeric vectors that encode the semantic meaning of data (text, images, audio). Semantically similar items end up close in the vector space. Embeddings are the foundation of semantic search, RAG (vector databases), and clustering. They are not backups, templates, or instances.

Context Window

1 question

Question 22
What is the "context window" of a Large Language Model?
Correct answer: B. The maximum number of tokens the model can process in a single call (input + output)
The context window is the maximum number of tokens an LLM can process in a single call — including input prompt and generated response. Modern models range from a few thousand (4K, 8K) to millions of tokens. When this limit is exceeded, the model loses access to the initial text. It has nothing to do with UI, concurrent users, or training time.

Hallucinations

1 question

Question 23
What are "hallucinations" in generative AI models?
Correct answer: B. Responses generated by the model that look plausible but are factually wrong or made up
Hallucinations are convincingly plausible responses generated by the model that are factually wrong or made up (dates, names, facts). They happen because LLMs are trained to generate plausible text, not to verify facts. Mitigations include RAG (providing sources), Guardrails, and human validation. They are not hardware bugs, rate-limits, or images.

Multimodal Models

1 question

Question 24
What characterizes a "multimodal" AI model?
Correct answer: B. A model capable of processing and/or generating multiple data types (e.g., text, image, audio, video)
Multimodal models accept and/or generate multiple data types: text, image, audio, video. Examples include Claude 3, GPT-4o, Gemini, Anthropic Vision. "Multilingual" (multiple languages) is different from multimodal. Multi-region and parallelism are infrastructure aspects, not model capabilities.

Knowledge Bases

1 question

Question 25
Which Amazon Bedrock feature allows you to implement RAG in a fully managed way, connecting your documents to a foundation model?
Correct answer: A. Bedrock Knowledge Bases
Bedrock Knowledge Bases is the managed RAG feature: you point to a data source (S3, SharePoint, Salesforce, etc.), and Bedrock handles indexing, embedding creation, vector store, and retrieval at query time. Pricing Calculator estimates costs, Macie classifies sensitive data, and Lambda Layers are shared libraries — none implement RAG.

Agents (Bedrock)

1 question

Question 26
What are Agents for Amazon Bedrock used for?
Correct answer: B. Allow foundation models to execute actions on external systems (APIs, databases) through multi-step reasoning
Agents for Amazon Bedrock allow foundation models to execute complex tasks by calling APIs, querying databases, and making multi-step decisions (function calling + reasoning). You define available actions and the agent decides when and how to call them. It does not replace IAM, host models, or configure load balancers.

Guardrails

1 question

Question 27
Which Amazon Bedrock feature allows you to define filters to block harmful content, denied topics, and personal information (PII) in prompts and responses?
Correct answer: B. Amazon Bedrock Guardrails
Bedrock Guardrails lets you configure customizable filters at the application level: block harmful content, denied topics (subjects the model must not address), and PII (emails, IDs, etc.). It works on both input (prompt) and output (response). WAF protects general web apps, Macie classifies data in S3, and Shield protects against DDoS.

SageMaker JumpStart

1 question

Question 28
Which Amazon SageMaker feature provides a catalog of foundation models and pre-trained models that can be deployed in just a few clicks?
Correct answer: B. SageMaker JumpStart
SageMaker JumpStart is a model hub with FMs (Llama, Falcon, Stable Diffusion, etc.) and pre-trained task-specific models, ready for deployment or fine-tuning. Pipelines orchestrates ML flows, Clarify measures bias, and Model Monitor detects model drift in production.

Bedrock Pricing

2 questions

Question 29
Which billing model does Amazon Bedrock use by default (on-demand) when invoking a foundation model?
Correct answer: B. Billing per number of input and output tokens processed
Bedrock on-demand bills by the number of tokens processed — separately input tokens (prompt) and output tokens (response), with different prices per model. It is not per-instance billing (no server to manage) nor a fixed subscription. Provisioned Throughput is an alternative for heavy usage, but not the default.
Question 43
When a company has constant high-volume use of a Bedrock foundation model, which billing model offers reserved capacity with consistent latency?
Correct answer: B. Provisioned Throughput (reserved capacity with commitment)
Bedrock Provisioned Throughput lets you reserve inference capacity (in "model units") with a 1- or 6-month commitment, ensuring consistent latency and a discount versus on-demand for heavy usage. On-Demand bills per token but may throttle during peaks, Free Tier is for experimentation, and Spot Instances are EC2 (not Bedrock).

Fairness

1 question

Question 30
What does "fairness" mean when building responsible AI systems?
Correct answer: B. The model must produce consistent, non-discriminatory results across different demographic groups
AI fairness means the system should produce equitable, non-discriminatory results across different groups (gender, race, age, region, etc.), avoiding perpetuating or amplifying historical inequalities. It has no relation to response time, open-source licensing, or pricing model.

Explainability

1 question

Question 31
Which capability describes the degree to which humans can understand why an AI model made a specific decision?
Correct answer: B. Explainability / interpretability
Explainability is the ability to make AI model decisions understandable — fundamental in regulated domains (healthcare, finance, justice). It includes techniques such as SHAP, LIME, attention maps, and tools like SageMaker Clarify. Latency and throughput are performance metrics, and quantization is a model compression technique.

Bias Detection

1 question

Question 32
Which Amazon SageMaker feature helps detect and explain bias and generate explainability reports for ML models?
Correct answer: A. SageMaker Clarify
SageMaker Clarify is the AWS tool dedicated to bias analysis (during and after training) and explainability reports (SHAP values, feature importance). Pipelines orchestrates ML flows, Studio is the ML IDE, and Endpoints host models for inference — none focus on bias/explainability.

Data Privacy

1 question

Question 33
A company wants to use Bedrock foundation models to process confidential data without exposing the content to other customers or to the model providers. Which Bedrock characteristic meets this need?
Correct answer: A. Data stays isolated within the customer account, never leaves the AWS tenant, and is not used for training
Bedrock guarantees data isolation within the customer account — prompt and response content is not shared between customers, with model providers, or used for retraining foundation models. This isolation is central to confidential use cases. The other options describe the opposite of what the service guarantees.

Compliance

2 questions

Question 34
Which AWS service can be used to access compliance reports (SOC, ISO, HIPAA, GDPR) applicable to AWS AI services such as Bedrock and SageMaker?
Correct answer: A. AWS Artifact
AWS Artifact is the AWS self-service portal for compliance reports, providing access to documents such as SOC reports, ISO certifications, PCI DSS, HIPAA BAA, etc. It covers all AWS services, including AI ones. Rekognition is a computer vision service, Lambda is serverless compute, and Knowledge Bases is managed RAG — none provide compliance reports.
Question 50
A health-tech startup wants to use Bedrock foundation models to process patient data in the US. Which compliance aspect is most relevant?
Correct answer: A. HIPAA (Health Insurance Portability and Accountability Act)
HIPAA regulates the privacy and security of health information (PHI - Protected Health Information) in the US. Bedrock is HIPAA-eligible with a signed BAA (Business Associate Agreement), per the AWS HIPAA Eligible Services list. ISO 9001 is quality management, PCI DSS protects credit card data, and SOX is corporate financial compliance — none specific to healthcare.

AWS AI Services

2 questions

Question 35
Which AWS service is a pre-trained API that extracts entities, sentiment, and topics from text without requiring a custom model?
Correct answer: A. Amazon Comprehend
Amazon Comprehend is the AWS managed NLP service with pre-trained APIs for sentiment analysis, entity extraction, language detection, classification, and topic modeling. It does not require training. SageMaker is the general ML platform, Glue is ETL, and Athena is SQL queries on S3.
Question 37
Which AWS service would you use to extract text from images and scanned PDFs (OCR)?
Correct answer: A. Amazon Textract
Amazon Textract uses ML to extract text, forms, and tables from scanned documents (PDFs, images). Polly is text-to-speech (generates audio from text), Translate is automatic translation, and Lex is for building chatbots — none of them perform OCR.

Overfitting

1 question

Question 36
In machine learning, what characterizes a model with "overfitting"?
Correct answer: B. The model performs well on training data but poorly on new data (it does not generalize)
Overfitting happens when the model memorizes specific patterns from training data (including noise) instead of learning generalizable ones. Result: high accuracy on training and low on new data. Underfitting (model too simple) has poor performance on both. Speed is not the criterion.

Inference Parameters

2 questions

Question 38
What does the "temperature" parameter control in Large Language Model text generation?
Correct answer: B. The level of randomness/creativity of the response — low values produce deterministic outputs, high values produce more varied outputs
Temperature is a parameter between 0 and 1 (some models go up to 2) that controls sampling randomness in generation: temperature=0 produces the most likely (deterministic) output, high temperature (~0.9) produces more creative/varied responses. It is not related to hardware, cost, or response length.
Question 39
Which inference parameter limits the set of candidate tokens during generation to the smallest group whose cumulative probability reaches a threshold (e.g., 0.9), providing additional control over randomness together with temperature?
Correct answer: B. Top-P (nucleus sampling)
Top-P (nucleus sampling) restricts sampling to the most probable tokens whose probability sum reaches the configured value (e.g., 0.9 includes tokens until covering 90% of the mass). It combines with temperature to control diversity. Max Tokens limits response length, Stop Sequences stop generation when a string appears, and Frequency Penalty discourages repetitions.

GenAI Use Cases

1 question

Question 40
What are typical use cases of generative AI? (Choose two.)
Correct answer: A. Generating text, code, and automatic summaries | C. Creating images and digital art from textual descriptions
Generative AI is used to create content: text (summaries, emails, articles), code (assistants such as Q Developer), images/art (text-to-image like Stable Diffusion), audio, and video. Network traffic analysis (CloudWatch, GuardDuty), hardware replacement (compute), and billing optimization (Cost Explorer, Trusted Advisor) are NOT GenAI use cases.

GenAI Limitations

1 question

Question 41
What are known limitations of current generative AI models?
Correct answer: B. The models can produce hallucinations, have a knowledge cutoff date, and reflect training data biases
Common limitations of LLMs/foundation models include: hallucinations (plausible but incorrect answers), cutoff date (no knowledge of events after training), bias inherited from data, lack of deep reasoning on complex problems, and computational cost. They do NOT always provide sources (RAG is needed for that) nor fail at multilingual tasks.

Vector Database

1 question

Question 44
To implement RAG, it is common to store embeddings in a database specialized in vector similarity search. Which AWS service can be used as a native vector store for Bedrock Knowledge Bases?
Correct answer: A. Amazon OpenSearch Serverless
Amazon OpenSearch Serverless is one of the native vector stores supported by Bedrock Knowledge Bases (alongside Aurora PostgreSQL with pgvector and Pinecone). Athena queries data in S3 via SQL, Redshift is a data warehouse, and Glue is ETL — none are vector databases.

PartyRock

1 question

Question 45
Which AWS service lets you quickly build and share generative AI apps without code, using Bedrock foundation models?
Correct answer: A. AWS PartyRock
PartyRock is a free AWS playground to build generative AI apps visually and without code, using Bedrock models under the hood. Lambda is serverless compute, CloudFront is a CDN, and SNS is pub/sub — none of them is a no-code AI platform.

Robustness

1 question

Question 46
Which responsible AI practice helps ensure a model continues to work correctly when receiving adversarial or out-of-distribution inputs?
Correct answer: B. Robustness and adversarial testing
Robustness is the model’s ability to maintain performance on atypical, noisy, or adversarial inputs (intentionally designed to mislead). Adversarial testing and continuous validation are pillars of responsible AI. Latency is a performance metric, and changing model size does not guarantee robustness on its own.

Ground Truth

1 question

Question 47
Which AWS service helps create high-quality labeled datasets to train ML models, offering management of human annotation workflows?
Correct answer: A. Amazon SageMaker Ground Truth
SageMaker Ground Truth is the AWS service for managing data labeling workflows with support for human annotators (internal, Mechanical Turk, or vendors), reducing cost and time. Quality labeled data is fundamental to mitigate bias and produce responsible AI. DataSync transfers files, CloudWatch monitors, and DataBrew prepares data without code (but does not label).

VPC Endpoints

1 question

Question 48
How can a company access Amazon Bedrock from a private VPC without traffic going through the public internet?
Correct answer: B. Use VPC Endpoints (PrivateLink) for Bedrock
VPC Endpoints (powered by AWS PrivateLink) allow access to Bedrock through a private interface inside the VPC, without traffic over the public internet. This increases security and compliance. VPN and Direct Connect connect on-prem to AWS in general but do not specifically isolate traffic to the Bedrock service. IPv6 does not serve this purpose.

Logging

1 question

Question 49
For auditing and compliance purposes, how can a company log all invocations (prompts and responses) made to Bedrock foundation models?
Correct answer: A. Enable Model Invocation Logging on Bedrock, sending logs to CloudWatch Logs or S3
Bedrock offers the "Model Invocation Logging" feature that, when enabled, records prompts, responses, and metadata in CloudWatch Logs and/or S3 — useful for auditing, debugging, and compliance. Logs stay in the customer’s account (not automatic). A custom proxy would be possible but unnecessary.

Other AWS mock exams

Practice for other AWS certifications with the same study and exam modes.

CLF-C02

AWS Cloud Practitioner

Foundational · 65 questions

MLA-C01

AWS ML Engineer Associate

Associate · 65 questions

See all AWS mock exams →

Frequently Asked Questions

0/5

✓How many questions are in the AWS AI Practitioner mock exam?

The mock exam includes 50 practical AIF-C01-style questions covering the 5 official domains, each with a detailed explanation to support your learning.

✓Are the questions based on the AIF-C01 exam?

Yes. The content is organized to reflect the 5 official domains of the AWS Certified AI Practitioner (AIF-C01): AI/ML fundamentals, generative AI, foundation models, responsible AI, and AI security & governance.

✓Do I get an explanation after answering?

Yes. After selecting an option, the mock exam shows the explanation behind the correct answer and why the other options are wrong.

✓What is the passing score on the official exam?

AIF-C01 uses a compensatory scoring model from 100 to 1000, with a minimum passing score of 700 (~70%). The mock exam adopts the same 70% threshold.

✓Is this mock exam free?

Yes. The full mock exam is free and available online without sign-up.

AWS CertifiedAIF-C01 Exam

Mock Exam AI Practitioner

Prepare for the AWS AI Practitioner (AIF-C01) exam with practical questions and detailed explanations.

Free

Detailed explanations

Updated 2026

This mock exam is an independent educational resource. It is not affiliated with, endorsed by, or sponsored by Amazon Web Services (AWS).

Practice by domain

How do you want to train?

Choose the mode that fits your current goal.

Study Mode

Feedback after every answer, free navigation, no timer.

✓Full explanation after each answer
✓Go back and review at any time
✓No time limit

Exam Mode

Realistic simulation of the actual exam with a countdown.

✓90 minutes to answer everything
✓No intermediate feedback
✓Score and answer key only at the end

Questions and commented answer key

See all questions with the correct answers and detailed explanations to reinforce your learning.

Questions

Topics

AI Definition

1 question

Question 1
What is Artificial Intelligence (AI) in its most basic definition?
Correct answer: B. Computers performing tasks that would normally require human intelligence
AI is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as pattern recognition, decision-making, and natural language processing. It is not just storage, specific hardware, or a programming language.

ML Types

1 question

Question 2
Which type of machine learning uses labeled data (with the correct answer) to train a model to make predictions?
Correct answer: A. Supervised learning
Supervised learning uses labeled data (input → expected output) to train classification or regression models. Unsupervised learning finds patterns in unlabeled data (clustering), reinforcement learning learns through trial and error with rewards, and deep learning is a technique (deep neural networks) applicable to any of these types.

ML Pipeline

1 question

Question 3
In a typical machine learning pipeline, what is the difference between the training and inference phases?
Correct answer: B. Training adjusts model parameters from data; inference uses the trained model to make predictions on new data
Training is the phase where the model learns by adjusting its parameters (weights) from historical data. Inference (also called prediction) is the phase where the trained model processes new data to generate results. The two phases may use different hardware, but the core distinction is that training creates/updates the model, while inference uses it.

AWS ML Services

1 question

Question 4
Which AWS service is a fully managed platform to develop, train, and deploy machine learning models at scale?
Correct answer: B. Amazon SageMaker
Amazon SageMaker is the AWS managed platform for the full ML lifecycle: development (Studio, notebooks), training, deploying models to endpoints, and monitoring. Lambda is serverless compute, Comprehend is a pre-trained NLP service for specific tasks, and Glue is data ETL.

GenAI Concepts

1 question

Question 5
What differentiates Generative AI from traditional AI?
Correct answer: B. Generative AI creates new content (text, image, code, audio) instead of just classifying or predicting
Generative AI creates new content (text, images, code, audio, video) based on patterns learned during training, while traditional AI focuses on classifying, predicting, or analyzing existing data. It uses ML models (especially foundation models), works with various data types, and can run on any platform.

Foundation Models

1 question

Question 6
What is a Foundation Model (FM) in the context of generative AI?
Correct answer: B. A large-scale model pre-trained on massive amounts of data that can be adapted for various downstream tasks
Foundation Models are AI models pre-trained on massive amounts of data (texts, images, etc.) that serve as a "foundation" to be adapted to multiple tasks via fine-tuning, RAG, or prompt engineering. Examples: Claude, Llama, Titan, Stable Diffusion. They are not instances, frameworks, or single-task models.

Tokens

1 question

Question 7
In the context of Large Language Models (LLMs), what is a token?
Correct answer: B. The basic unit of text the model processes, typically a word or part of a word
Tokens are the basic units LLMs process — they can be whole words, sub-words, or characters depending on the tokenizer. Usage cost and model context limits are measured in tokens. Not to be confused with authentication tokens or encryption keys.

Prompts

1 question

Question 8
What is a prompt in the context of generative AI?
Correct answer: B. The input (text, image, or other data) provided to the model to guide it in generating a response
A prompt is the input (usually text) the user provides to the model to guide its generation — it may include instructions, context, examples, and questions. Prompt quality directly affects response quality (hence the discipline of "prompt engineering"). It is not an instance, password, or model weight.

Amazon Bedrock

1 question

Question 9
Which AWS service provides managed access to foundation models from multiple providers (Anthropic, Meta, Cohere, Amazon, etc.) through a unified API?
Correct answer: B. Amazon Bedrock
Amazon Bedrock is the AWS managed service that provides access to foundation models from various providers via a single API, without having to manage infrastructure. SageMaker is a general ML platform, Lambda is serverless compute, and Comprehend is pre-trained NLP for specific tasks.

Prompt Engineering

2 questions

Question 10
Which prompt engineering technique involves providing the model with a few examples of (input, expected output) pairs before the actual question to guide the response style?
Correct answer: B. Few-shot prompting
Few-shot prompting provides the model with a few examples (typically 2-5) inside the prompt so it learns the expected pattern. Zero-shot is without examples (instruction only), fine-tuning retrains model weights with new data (not prompt engineering), and pre-training is the original foundation model creation phase.
Question 42
Which prompt engineering technique asks the model to "think step by step" before giving the final answer, improving performance on reasoning problems?
Correct answer: A. Chain-of-Thought (CoT) prompting
Chain-of-Thought (CoT) prompting instructs the model to explicitly reason step by step before the final answer. Particularly useful for math, logic, and planning problems. Zero-shot is instruction without examples, token streaming is incremental token delivery, and quantization is a model compression technique.

RAG

1 question

Question 11
Which technique allows a foundation model to answer using company-specific, up-to-date information without retraining the model?
Correct answer: B. Retrieval Augmented Generation (RAG)
RAG (Retrieval Augmented Generation) retrieves relevant documents from a knowledge base (typically a vector database such as Bedrock Knowledge Bases) and injects the content into the prompt before generation, letting the model answer using specific data without retraining. Pre-training creates the model from scratch, quantization reduces size, and tokenization splits text into tokens.

Fine-Tuning

1 question

Question 12
When does it make more sense to apply fine-tuning to a foundation model instead of using prompt engineering?
Correct answer: B. When you need to adapt the model to a very specific domain with its own vocabulary and have quality labeled data
Fine-tuning is appropriate when the use case requires deep adaptation to a specific domain (technical jargon, consistent format, very specific behavior) and quality training data exists. It is more expensive and time-consuming than prompt engineering, so it should only be used when prompt engineering is insufficient. For light tone or style adjustments, prompt engineering is enough.

Amazon Q

1 question

Question 13
Which AWS service is a generative AI assistant for work that can answer questions, generate content, and automate tasks, focused on enterprise environments?
Correct answer: A. Amazon Q
Amazon Q is the AWS generative AI assistant focused on enterprise use cases (Q Developer for code, Q Business for company knowledge). Lex is for building custom chatbots, Polly does text-to-speech, and Comprehend is pre-trained NLP for text analysis.

Bias

1 question

Question 14
What is "bias" in Artificial Intelligence systems?
Correct answer: B. Systematic tendencies that lead to unfair or inaccurate results for certain groups, typically inherited from training data
AI bias refers to systematic tendencies in the model that produce unequal or unfair results for specific groups (gender, race, age, etc.). It typically comes from training data that reflects historical inequalities or imbalanced sampling. Mitigating bias is a central pillar of responsible AI.

AI Service Cards

1 question

Question 15
What are AWS AI Service Cards?
Correct answer: B. Documents published by AWS that describe intended use cases, limitations, responsible design considerations, and best practices for each AI service
AWS AI Service Cards are transparency documents published by AWS for AI services (such as Rekognition Face Liveness, Bedrock FMs, etc.), describing intended use cases, limitations, responsible design considerations, and best practices. They serve as a guide for ethical service use.

Encryption

1 question

Question 16
How does Amazon Bedrock protect data that customers send to foundation models by default?
Correct answer: B. Data is encrypted in transit (TLS) and at rest, and is not used to train other customers’ models
Amazon Bedrock encrypts all data in transit (TLS) and at rest, keeps data isolated within the customer’s account, and does NOT use prompt or response content to train foundation models or share with third parties. This is a core data governance principle of Bedrock.

IAM

1 question

Question 17
Which AWS security practice should be used to grant an application access to Amazon SageMaker without exposing permanent credentials?
Correct answer: B. Use IAM Roles with temporary permissions
IAM Roles are the AWS recommended practice for granting access to services (including SageMaker) because they provide rotating temporary credentials and follow the principle of least privilege. Sharing root credentials, hardcoding, or disabling IAM directly violate AWS security best practices.

Deep Learning

1 question

Question 18
Which machine learning technique uses neural networks with multiple layers to learn complex data representations, serving as the foundation of many generative AI models?
Correct answer: C. Deep Learning (deep neural networks)
Deep Learning uses neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data. It is the foundation of modern AI models, including CNNs (vision), RNNs/Transformers (language), and foundation models. Linear regression, decision trees, and Naive Bayes are simpler ML techniques that do not involve multiple neural layers.

Classification vs Regression

1 question

Question 19
What is the difference between a classification model and a regression model in machine learning?
Correct answer: A. Classification predicts discrete categories (e.g., spam vs. not-spam); regression predicts continuous numeric values (e.g., price)
Classification predicts discrete categories (spam/not-spam, sick/healthy, cat/dog). Regression predicts continuous numeric values (price, temperature, age). The choice depends on the target variable type. It has nothing to do with speed or input type — both can use images, numbers, or text as inputs.

Model Evaluation

1 question

Question 20
In a binary classification problem with highly imbalanced data (e.g., fraud detection), which metric is generally more informative than simple accuracy?
Correct answer: B. F1-score (combination of precision and recall)
F1-score combines precision (correct predictions among predicted positives) and recall (actual positives detected), being more informative than accuracy on imbalanced datasets. In fraud detection, a model that always predicts "non-fraud" might achieve 99% accuracy but be useless. The other options (time, size, epochs) are operational details, not model quality metrics.

Embeddings

1 question

Question 21
What are embeddings (vector representations) in the context of generative AI?
Correct answer: B. Numeric vectors that represent the semantic meaning of texts, images, or other data, enabling similarity comparisons
Embeddings are high-dimensional numeric vectors that encode the semantic meaning of data (text, images, audio). Semantically similar items end up close in the vector space. Embeddings are the foundation of semantic search, RAG (vector databases), and clustering. They are not backups, templates, or instances.

Context Window

1 question

Question 22
What is the "context window" of a Large Language Model?
Correct answer: B. The maximum number of tokens the model can process in a single call (input + output)
The context window is the maximum number of tokens an LLM can process in a single call — including input prompt and generated response. Modern models range from a few thousand (4K, 8K) to millions of tokens. When this limit is exceeded, the model loses access to the initial text. It has nothing to do with UI, concurrent users, or training time.

Hallucinations

1 question

Question 23
What are "hallucinations" in generative AI models?
Correct answer: B. Responses generated by the model that look plausible but are factually wrong or made up
Hallucinations are convincingly plausible responses generated by the model that are factually wrong or made up (dates, names, facts). They happen because LLMs are trained to generate plausible text, not to verify facts. Mitigations include RAG (providing sources), Guardrails, and human validation. They are not hardware bugs, rate-limits, or images.

Multimodal Models

1 question

Question 24
What characterizes a "multimodal" AI model?
Correct answer: B. A model capable of processing and/or generating multiple data types (e.g., text, image, audio, video)
Multimodal models accept and/or generate multiple data types: text, image, audio, video. Examples include Claude 3, GPT-4o, Gemini, Anthropic Vision. "Multilingual" (multiple languages) is different from multimodal. Multi-region and parallelism are infrastructure aspects, not model capabilities.

Knowledge Bases

1 question

Question 25
Which Amazon Bedrock feature allows you to implement RAG in a fully managed way, connecting your documents to a foundation model?
Correct answer: A. Bedrock Knowledge Bases
Bedrock Knowledge Bases is the managed RAG feature: you point to a data source (S3, SharePoint, Salesforce, etc.), and Bedrock handles indexing, embedding creation, vector store, and retrieval at query time. Pricing Calculator estimates costs, Macie classifies sensitive data, and Lambda Layers are shared libraries — none implement RAG.

Agents (Bedrock)

1 question

Question 26
What are Agents for Amazon Bedrock used for?
Correct answer: B. Allow foundation models to execute actions on external systems (APIs, databases) through multi-step reasoning
Agents for Amazon Bedrock allow foundation models to execute complex tasks by calling APIs, querying databases, and making multi-step decisions (function calling + reasoning). You define available actions and the agent decides when and how to call them. It does not replace IAM, host models, or configure load balancers.

Guardrails

1 question

Question 27
Which Amazon Bedrock feature allows you to define filters to block harmful content, denied topics, and personal information (PII) in prompts and responses?
Correct answer: B. Amazon Bedrock Guardrails
Bedrock Guardrails lets you configure customizable filters at the application level: block harmful content, denied topics (subjects the model must not address), and PII (emails, IDs, etc.). It works on both input (prompt) and output (response). WAF protects general web apps, Macie classifies data in S3, and Shield protects against DDoS.

SageMaker JumpStart

1 question

Question 28
Which Amazon SageMaker feature provides a catalog of foundation models and pre-trained models that can be deployed in just a few clicks?
Correct answer: B. SageMaker JumpStart
SageMaker JumpStart is a model hub with FMs (Llama, Falcon, Stable Diffusion, etc.) and pre-trained task-specific models, ready for deployment or fine-tuning. Pipelines orchestrates ML flows, Clarify measures bias, and Model Monitor detects model drift in production.

Bedrock Pricing

2 questions

Question 29
Which billing model does Amazon Bedrock use by default (on-demand) when invoking a foundation model?
Correct answer: B. Billing per number of input and output tokens processed
Bedrock on-demand bills by the number of tokens processed — separately input tokens (prompt) and output tokens (response), with different prices per model. It is not per-instance billing (no server to manage) nor a fixed subscription. Provisioned Throughput is an alternative for heavy usage, but not the default.
Question 43
When a company has constant high-volume use of a Bedrock foundation model, which billing model offers reserved capacity with consistent latency?
Correct answer: B. Provisioned Throughput (reserved capacity with commitment)
Bedrock Provisioned Throughput lets you reserve inference capacity (in "model units") with a 1- or 6-month commitment, ensuring consistent latency and a discount versus on-demand for heavy usage. On-Demand bills per token but may throttle during peaks, Free Tier is for experimentation, and Spot Instances are EC2 (not Bedrock).

Fairness

1 question

Question 30
What does "fairness" mean when building responsible AI systems?
Correct answer: B. The model must produce consistent, non-discriminatory results across different demographic groups
AI fairness means the system should produce equitable, non-discriminatory results across different groups (gender, race, age, region, etc.), avoiding perpetuating or amplifying historical inequalities. It has no relation to response time, open-source licensing, or pricing model.

Explainability

1 question

Question 31
Which capability describes the degree to which humans can understand why an AI model made a specific decision?
Correct answer: B. Explainability / interpretability
Explainability is the ability to make AI model decisions understandable — fundamental in regulated domains (healthcare, finance, justice). It includes techniques such as SHAP, LIME, attention maps, and tools like SageMaker Clarify. Latency and throughput are performance metrics, and quantization is a model compression technique.

Bias Detection

1 question

Question 32
Which Amazon SageMaker feature helps detect and explain bias and generate explainability reports for ML models?
Correct answer: A. SageMaker Clarify
SageMaker Clarify is the AWS tool dedicated to bias analysis (during and after training) and explainability reports (SHAP values, feature importance). Pipelines orchestrates ML flows, Studio is the ML IDE, and Endpoints host models for inference — none focus on bias/explainability.

Data Privacy

1 question

Question 33
A company wants to use Bedrock foundation models to process confidential data without exposing the content to other customers or to the model providers. Which Bedrock characteristic meets this need?
Correct answer: A. Data stays isolated within the customer account, never leaves the AWS tenant, and is not used for training
Bedrock guarantees data isolation within the customer account — prompt and response content is not shared between customers, with model providers, or used for retraining foundation models. This isolation is central to confidential use cases. The other options describe the opposite of what the service guarantees.

Compliance

2 questions

Question 34
Which AWS service can be used to access compliance reports (SOC, ISO, HIPAA, GDPR) applicable to AWS AI services such as Bedrock and SageMaker?
Correct answer: A. AWS Artifact
AWS Artifact is the AWS self-service portal for compliance reports, providing access to documents such as SOC reports, ISO certifications, PCI DSS, HIPAA BAA, etc. It covers all AWS services, including AI ones. Rekognition is a computer vision service, Lambda is serverless compute, and Knowledge Bases is managed RAG — none provide compliance reports.
Question 50
A health-tech startup wants to use Bedrock foundation models to process patient data in the US. Which compliance aspect is most relevant?
Correct answer: A. HIPAA (Health Insurance Portability and Accountability Act)
HIPAA regulates the privacy and security of health information (PHI - Protected Health Information) in the US. Bedrock is HIPAA-eligible with a signed BAA (Business Associate Agreement), per the AWS HIPAA Eligible Services list. ISO 9001 is quality management, PCI DSS protects credit card data, and SOX is corporate financial compliance — none specific to healthcare.

AWS AI Services

2 questions

Question 35
Which AWS service is a pre-trained API that extracts entities, sentiment, and topics from text without requiring a custom model?
Correct answer: A. Amazon Comprehend
Amazon Comprehend is the AWS managed NLP service with pre-trained APIs for sentiment analysis, entity extraction, language detection, classification, and topic modeling. It does not require training. SageMaker is the general ML platform, Glue is ETL, and Athena is SQL queries on S3.
Question 37
Which AWS service would you use to extract text from images and scanned PDFs (OCR)?
Correct answer: A. Amazon Textract
Amazon Textract uses ML to extract text, forms, and tables from scanned documents (PDFs, images). Polly is text-to-speech (generates audio from text), Translate is automatic translation, and Lex is for building chatbots — none of them perform OCR.

Overfitting

1 question

Question 36
In machine learning, what characterizes a model with "overfitting"?
Correct answer: B. The model performs well on training data but poorly on new data (it does not generalize)
Overfitting happens when the model memorizes specific patterns from training data (including noise) instead of learning generalizable ones. Result: high accuracy on training and low on new data. Underfitting (model too simple) has poor performance on both. Speed is not the criterion.

Inference Parameters

2 questions

Question 38
What does the "temperature" parameter control in Large Language Model text generation?
Correct answer: B. The level of randomness/creativity of the response — low values produce deterministic outputs, high values produce more varied outputs
Temperature is a parameter between 0 and 1 (some models go up to 2) that controls sampling randomness in generation: temperature=0 produces the most likely (deterministic) output, high temperature (~0.9) produces more creative/varied responses. It is not related to hardware, cost, or response length.
Question 39
Which inference parameter limits the set of candidate tokens during generation to the smallest group whose cumulative probability reaches a threshold (e.g., 0.9), providing additional control over randomness together with temperature?
Correct answer: B. Top-P (nucleus sampling)
Top-P (nucleus sampling) restricts sampling to the most probable tokens whose probability sum reaches the configured value (e.g., 0.9 includes tokens until covering 90% of the mass). It combines with temperature to control diversity. Max Tokens limits response length, Stop Sequences stop generation when a string appears, and Frequency Penalty discourages repetitions.

GenAI Use Cases

1 question

Question 40
What are typical use cases of generative AI? (Choose two.)
Correct answer: A. Generating text, code, and automatic summaries | C. Creating images and digital art from textual descriptions
Generative AI is used to create content: text (summaries, emails, articles), code (assistants such as Q Developer), images/art (text-to-image like Stable Diffusion), audio, and video. Network traffic analysis (CloudWatch, GuardDuty), hardware replacement (compute), and billing optimization (Cost Explorer, Trusted Advisor) are NOT GenAI use cases.

GenAI Limitations

1 question

Question 41
What are known limitations of current generative AI models?
Correct answer: B. The models can produce hallucinations, have a knowledge cutoff date, and reflect training data biases
Common limitations of LLMs/foundation models include: hallucinations (plausible but incorrect answers), cutoff date (no knowledge of events after training), bias inherited from data, lack of deep reasoning on complex problems, and computational cost. They do NOT always provide sources (RAG is needed for that) nor fail at multilingual tasks.

Vector Database

1 question

Question 44
To implement RAG, it is common to store embeddings in a database specialized in vector similarity search. Which AWS service can be used as a native vector store for Bedrock Knowledge Bases?
Correct answer: A. Amazon OpenSearch Serverless
Amazon OpenSearch Serverless is one of the native vector stores supported by Bedrock Knowledge Bases (alongside Aurora PostgreSQL with pgvector and Pinecone). Athena queries data in S3 via SQL, Redshift is a data warehouse, and Glue is ETL — none are vector databases.

PartyRock

1 question

Question 45
Which AWS service lets you quickly build and share generative AI apps without code, using Bedrock foundation models?
Correct answer: A. AWS PartyRock
PartyRock is a free AWS playground to build generative AI apps visually and without code, using Bedrock models under the hood. Lambda is serverless compute, CloudFront is a CDN, and SNS is pub/sub — none of them is a no-code AI platform.

Robustness

1 question

Question 46
Which responsible AI practice helps ensure a model continues to work correctly when receiving adversarial or out-of-distribution inputs?
Correct answer: B. Robustness and adversarial testing
Robustness is the model’s ability to maintain performance on atypical, noisy, or adversarial inputs (intentionally designed to mislead). Adversarial testing and continuous validation are pillars of responsible AI. Latency is a performance metric, and changing model size does not guarantee robustness on its own.

Ground Truth

1 question

Question 47
Which AWS service helps create high-quality labeled datasets to train ML models, offering management of human annotation workflows?
Correct answer: A. Amazon SageMaker Ground Truth
SageMaker Ground Truth is the AWS service for managing data labeling workflows with support for human annotators (internal, Mechanical Turk, or vendors), reducing cost and time. Quality labeled data is fundamental to mitigate bias and produce responsible AI. DataSync transfers files, CloudWatch monitors, and DataBrew prepares data without code (but does not label).

VPC Endpoints

1 question

Question 48
How can a company access Amazon Bedrock from a private VPC without traffic going through the public internet?
Correct answer: B. Use VPC Endpoints (PrivateLink) for Bedrock
VPC Endpoints (powered by AWS PrivateLink) allow access to Bedrock through a private interface inside the VPC, without traffic over the public internet. This increases security and compliance. VPN and Direct Connect connect on-prem to AWS in general but do not specifically isolate traffic to the Bedrock service. IPv6 does not serve this purpose.

Logging

1 question

Question 49
For auditing and compliance purposes, how can a company log all invocations (prompts and responses) made to Bedrock foundation models?
Correct answer: A. Enable Model Invocation Logging on Bedrock, sending logs to CloudWatch Logs or S3
Bedrock offers the "Model Invocation Logging" feature that, when enabled, records prompts, responses, and metadata in CloudWatch Logs and/or S3 — useful for auditing, debugging, and compliance. Logs stay in the customer’s account (not automatic). A custom proxy would be possible but unnecessary.

Other AWS mock exams

Practice for other AWS certifications with the same study and exam modes.

CLF-C02

AWS Cloud Practitioner

Foundational · 65 questions

MLA-C01

AWS ML Engineer Associate

Associate · 65 questions

See all AWS mock exams →

Frequently Asked Questions

0/5

✓How many questions are in the AWS AI Practitioner mock exam?

The mock exam includes 50 practical AIF-C01-style questions covering the 5 official domains, each with a detailed explanation to support your learning.

✓Are the questions based on the AIF-C01 exam?

✓Do I get an explanation after answering?

Yes. After selecting an option, the mock exam shows the explanation behind the correct answer and why the other options are wrong.

✓What is the passing score on the official exam?

AIF-C01 uses a compensatory scoring model from 100 to 1000, with a minimum passing score of 700 (~70%). The mock exam adopts the same 70% threshold.

✓Is this mock exam free?

Yes. The full mock exam is free and available online without sign-up.

Practice by domain

How do you want to train?

Study Mode

Exam Mode

Questions and commented answer key

AI Definition

Question 1

ML Types

Question 2

ML Pipeline

Question 3

AWS ML Services

Question 4

GenAI Concepts

Question 5

Foundation Models

Question 6

Tokens

Question 7

Prompts

Question 8

Amazon Bedrock

Question 9

Prompt Engineering

Question 10

Question 42

RAG

Question 11

Fine-Tuning

Question 12

Amazon Q

Question 13

Bias

Question 14

AI Service Cards

Question 15

Encryption

Question 16

IAM

Question 17

Deep Learning

Question 18

Classification vs Regression

Question 19

Model Evaluation

Question 20

Embeddings

Question 21

Context Window

Question 22

Hallucinations

Question 23

Multimodal Models

Question 24

Knowledge Bases

Question 25

Agents (Bedrock)

Question 26

Guardrails

Question 27

SageMaker JumpStart

Question 28

Bedrock Pricing

Question 29

Question 43

Fairness

Question 30

Explainability

Question 31

Bias Detection

Question 32

Data Privacy

Question 33

Compliance

Question 34

Question 50

AWS AI Services

Question 35

Question 37

Overfitting