AWS AIF-C01 · Updated 2026
Free AWS AI Practitioner Practice Exam (AIF-C01)
Full-length AIF-C01 mock exams with the same number of questions and time as the real exam. AI and ML basics, generative AI, foundation models, responsible AI and governance questions, answers explained, and your score by domain.
1. Choose a set
2. Choose a mode
65 questions · 1 hour 30 min · same as the real exam
You have an unfinished attempt on this set.
Questions 1–10Page 1 of 7
🏆
Question 1Foundation Model Applications
A company wants to deploy an open-source model on SageMaker endpoints it controls, choosing the instance types and network settings itself. Which option fits BEST?
Correct answer: A. Amazon SageMaker JumpStart
SageMaker JumpStart offers pretrained and open models that you can deploy to SageMaker endpoints with your own choice of instances and settings. Bedrock on-demand is serverless, so you do not pick or manage the underlying instances.
SageMaker JumpStart offers pretrained and open models that you can deploy to SageMaker endpoints with your own choice of instances and settings. Bedrock on-demand is serverless, so you do not pick or manage the underlying instances.
Question 2Generative AI Fundamentals
A developer notices that a 100-word reply was billed as about 130 tokens. Why?
Correct answer: C. Tokens can be word pieces and punctuation
Models split text into tokens, which can be whole words, parts of words, or punctuation, so token counts usually exceed word counts. The count depends on the model's tokenizer, not on a fixed fee or sentence count.
Models split text into tokens, which can be whole words, parts of words, or punctuation, so token counts usually exceed word counts. The count depends on the model's tokenizer, not on a fixed fee or sentence count.
Question 3AI & ML Fundamentals
A warehouse company wants a robot to learn the fastest route through its aisles. The robot gets a reward for quick deliveries and a penalty for collisions. Which type of learning is this?
Correct answer: D. Reinforcement learning
Reinforcement learning trains an agent through trial and error using rewards and penalties from its environment. Supervised learning needs a labeled correct answer for each example, which this setup does not provide.
Reinforcement learning trains an agent through trial and error using rewards and penalties from its environment. Supervised learning needs a labeled correct answer for each example, which this setup does not provide.
Question 4Responsible AI
A bank finds its model approves loans at very different rates for applicants of different ethnic groups who have similar finances. Which responsible AI dimension is MOST at risk?
Correct answer: D. Fairness
Fairness means a model's outcomes do not unjustly favor or harm particular groups, and different approval rates for similar applicants signal a fairness problem. Latency and scalability are performance concerns unrelated to equitable outcomes.
Fairness means a model's outcomes do not unjustly favor or harm particular groups, and different approval rates for similar applicants signal a fairness problem. Latency and scalability are performance concerns unrelated to equitable outcomes.
Question 5Security & Governance
A SageMaker training job fails with an access denied error when reading training data from an S3 bucket. What is the BEST fix?
Correct answer: B. Give the job's execution role S3 read access
SageMaker jobs act through an IAM execution role, so that role needs permission to read the bucket, following least privilege. Making the bucket public exposes the data to anyone and breaks security best practice.
SageMaker jobs act through an IAM execution role, so that role needs permission to read the bucket, following least privilege. Making the bucket public exposes the data to anyone and breaks security best practice.
Question 6Foundation Model Applications
A bank wants every chatbot reply to use a formal tone and refuse off-topic requests, no matter what users type. Where should these standing instructions go?
Correct answer: B. In the system prompt
The system prompt sets the role, tone, and rules the model should follow for the whole conversation. Putting rules only in user messages relies on each user and is easy to forget or override.
The system prompt sets the role, tone, and rules the model should follow for the whole conversation. Putting rules only in user messages relies on each user and is easy to forget or override.
Question 7Generative AI Fundamentals
A copywriting tool should let the model choose only from the smallest set of likely next tokens whose combined probability reaches 90%. Which parameter controls this?
Correct answer: D. Top-p
Top-p, also called nucleus sampling, limits choices to the most likely tokens whose cumulative probability reaches the set value, such as 0.9. Temperature reshapes how random the selection is but does not set a probability cutoff.
Top-p, also called nucleus sampling, limits choices to the most likely tokens whose cumulative probability reaches the set value, such as 0.9. Temperature reshapes how random the selection is but does not set a probability cutoff.
Question 8AI & ML Fundamentals
A bakery chain has recorded daily sales for each store over three years and wants to predict next month's demand. How should this data be described?
Correct answer: C. Time series data
Time series data is a sequence of values recorded at regular time intervals, which is exactly what daily sales per store are. Unstructured data refers to content such as free text, audio, or images that lacks a fixed format.
Time series data is a sequence of values recorded at regular time intervals, which is exactly what daily sales per store are. Unstructured data refers to content such as free text, audio, or images that lacks a fixed format.
Question 9Foundation Model Applications
A startup wants to improve answer quality for a new feature at the lowest cost and with no model training. What should it try FIRST?
Correct answer: B. Prompt engineering
Prompt engineering changes only the input to the model, so it is fast and cheap and needs no training data or compute jobs. Fine-tuning and continued pre-training cost more and should come after prompt improvements fall short.
Prompt engineering changes only the input to the model, so it is fast and cheap and needs no training data or compute jobs. Fine-tuning and continued pre-training cost more and should come after prompt improvements fall short.
Question 10Generative AI Fundamentals
A chatbot forgets what the user said two messages ago. The developer sends only the newest message on each API call. What is the cause?
Correct answer: C. The model is stateless and needs prior turns sent
Foundation model APIs do not remember previous calls, so the app must send the conversation history with each request, and that history counts toward the context window. Temperature affects randomness, not memory of earlier turns.
Foundation model APIs do not remember previous calls, so the app must send the conversation history with each request, and that history counts toward the context window. Temperature affects randomness, not memory of earlier turns.
Question 11Responsible AI
A clinic's chatbot must automatically mask phone numbers and email addresses in responses before users see them. Which Bedrock Guardrails feature fits?
Correct answer: A. Sensitive information filters
Sensitive information filters detect personally identifiable information such as phone numbers and emails and can block or mask it. Denied topics block whole subjects of conversation and do not redact specific data values.
Sensitive information filters detect personally identifiable information such as phone numbers and emails and can block or mask it. Denied topics block whole subjects of conversation and do not redact specific data values.
Question 12Security & Governance
A compliance officer needs to keep the full text of prompts and model responses for every Bedrock call for later review. What should the team enable?
Correct answer: B. Bedrock model invocation logging
Model invocation logging captures request and response data for Bedrock calls and delivers it to Amazon S3 or CloudWatch Logs. CloudTrail records who made the API call and when, but not the full prompt and response text.
Model invocation logging captures request and response data for Bedrock calls and delivers it to Amazon S3 or CloudWatch Logs. CloudTrail records who made the API call and when, but not the full prompt and response text.
Question 13Foundation Model Applications
A marketer prompts a model with 'Write about our new shoes' and gets bland, generic text. What is the BEST way to improve the prompt?
Correct answer: A. Add the audience, key features, tone, and length
Clear prompts that state the audience, key details, desired tone, and length give the model what it needs to write specific, useful text. Simply repeating a vague prompt will keep producing generic results.
Clear prompts that state the audience, key details, desired tone, and length give the model what it needs to write specific, useful text. Simply repeating a vague prompt will keep producing generic results.
Question 14AI & ML Fundamentals
A model performs poorly on both its training data and its test data. What is the MOST likely cause and fix?
Correct answer: C. Underfitting; add features or model complexity
Weak results on both training and test data mean the model is too simple to capture the patterns, which is underfitting. Adding regularization is a fix for overfitting and would make an underfit model even simpler.
Weak results on both training and test data mean the model is too simple to capture the patterns, which is underfitting. Adding regularization is a fix for overfitting and would make an underfit model even simpler.
Question 15Generative AI Fundamentals
A research team studies an architecture in which a generator creates fake images and a discriminator tries to tell them apart from real ones. Which architecture is this?
Correct answer: D. Generative adversarial network
A generative adversarial network trains two networks against each other: a generator that creates samples and a discriminator that judges whether they are real. Diffusion models instead learn to remove noise step by step and do not use a discriminator.
A generative adversarial network trains two networks against each other: a generator that creates samples and a discriminator that judges whether they are real. Diffusion models instead learn to remove noise step by step and do not use a discriminator.
Question 16Foundation Model Applications
A company already runs Amazon Aurora PostgreSQL and wants to store embeddings for a knowledge base in that same database engine. What does it need?
Correct answer: D. The pgvector extension
The pgvector extension adds vector storage and similarity search to PostgreSQL, and Aurora PostgreSQL with pgvector can serve as a Bedrock knowledge base vector store. Amazon Timestream is a time series database and does not provide vector search.
The pgvector extension adds vector storage and similarity search to PostgreSQL, and Aurora PostgreSQL with pgvector can serve as a Bedrock knowledge base vector store. Amazon Timestream is a time series database and does not provide vector search.
Question 17AI & ML Fundamentals
A ride-sharing app must estimate the fare within milliseconds while the rider waits on the booking screen. Which inference option fits BEST?
Correct answer: A. A real-time inference endpoint
Real-time inference endpoints return predictions with low latency for each request, which matches a user waiting on screen. Batch transform processes large datasets offline and cannot answer individual requests instantly.
Real-time inference endpoints return predictions with low latency for each request, which matches a user waiting on screen. Batch transform processes large datasets offline and cannot answer individual requests instantly.
Question 18Responsible AI
A retrieval-based app sometimes gives answers that are not supported by the retrieved documents. The team wants Guardrails to detect and filter these ungrounded responses. Which feature fits?
Correct answer: B. Contextual grounding checks
Contextual grounding checks score whether a response is grounded in the source content and relevant to the query, and they can block answers below a threshold. Word filters block specific words or phrases and cannot judge whether a claim matches the sources.
Contextual grounding checks score whether a response is grounded in the source content and relevant to the query, and they can block answers below a threshold. Word filters block specific words or phrases and cannot judge whether a claim matches the sources.
Question 19Generative AI Fundamentals
How does a large language model relate to a foundation model?
Correct answer: D. An LLM is a language-focused foundation model
Foundation models are large pretrained models that can be adapted to many tasks, and LLMs are the subset focused on understanding and generating text. Some foundation models handle images or several modalities rather than only language.
Foundation models are large pretrained models that can be adapted to many tasks, and LLMs are the subset focused on understanding and generating text. Some foundation models handle images or several modalities rather than only language.
Question 20Security & Governance
A company wants to be alerted if an S3 bucket holding ML training data is ever changed to allow public access or has encryption turned off. Which service evaluates resource settings against such rules?
Correct answer: C. AWS Config
AWS Config records resource configurations and evaluates them against rules, flagging noncompliant changes such as public access or missing encryption. AWS Artifact only provides AWS's compliance reports and agreements.
AWS Config records resource configurations and evaluates them against rules, flagging noncompliant changes such as public access or missing encryption. AWS Artifact only provides AWS's compliance reports and agreements.
Question 21Foundation Model Applications
A parts supplier's retrieval app often misses documents when users search exact part numbers such as XR-4410, because pure semantic search treats the codes loosely. Which retrieval change helps MOST?
Correct answer: A. Hybrid keyword and semantic search
Hybrid search blends keyword matching with vector similarity, so exact terms like part numbers are found while meaning-based matches still work. Changing temperature affects how the model writes, not which documents are retrieved.
Hybrid search blends keyword matching with vector similarity, so exact terms like part numbers are found while meaning-based matches still work. Changing temperature affects how the model writes, not which documents are retrieved.
Question 22Generative AI Fundamentals
During pretraining, how does a foundation model usually learn from huge amounts of text without people labeling each example?
Correct answer: C. By predicting hidden or next words in text
Foundation models are pretrained with self-supervised learning, where the text itself provides the target, such as the next word in a sentence. This avoids the cost of human labeling, which would be impossible at this scale.
Foundation models are pretrained with self-supervised learning, where the text itself provides the target, such as the next word in a sentence. This avoids the cost of human labeling, which would be impossible at this scale.
Question 23AI & ML Fundamentals
An email provider is tuning a spam filter. Users are very upset when a real email lands in the spam folder. Which metric matters MOST?
Correct answer: A. Precision
Precision measures how many emails flagged as spam are truly spam, so high precision means few real emails are wrongly flagged. Recall focuses on catching every spam message, which can raise false positives.
Precision measures how many emails flagged as spam are truly spam, so high precision means few real emails are wrongly flagged. Recall focuses on catching every spam message, which can raise false positives.
Question 24Foundation Model Applications
An evaluator wants a metric that rewards a generated answer for matching the meaning of a reference answer, even when it uses different words. Which metric fits BEST?
Correct answer: B. BERTScore
BERTScore compares contextual embeddings of the generated and reference text, so it captures semantic similarity beyond exact word overlap. BLEU counts matching word sequences and penalizes valid paraphrases.
BERTScore compares contextual embeddings of the generated and reference text, so it captures semantic similarity beyond exact word overlap. BLEU counts matching word sequences and penalizes valid paraphrases.
Question 25Responsible AI
A shipping company's address model works well on clean input but fails badly when addresses contain typos or extra spaces. Which responsible AI dimension needs work?
Correct answer: A. Robustness
Robustness is the ability of a model to keep working correctly with noisy, unexpected, or slightly changed inputs. Transparency is about being open regarding how AI is used, which does not fix failures on messy data.
Robustness is the ability of a model to keep working correctly with noisy, unexpected, or slightly changed inputs. Transparency is about being open regarding how AI is used, which does not fix failures on messy data.
Question 26Security & Governance
A bank must continuously collect evidence that its generative AI workloads follow internal and industry controls, ready for audits. Which service fits BEST?
Correct answer: C. AWS Audit Manager
AWS Audit Manager automates evidence collection against control frameworks, including a prebuilt framework for generative AI best practices. AWS Artifact gives access to AWS's own audit reports but does not gather evidence about your workloads.
AWS Audit Manager automates evidence collection against control frameworks, including a prebuilt framework for generative AI best practices. AWS Artifact gives access to AWS's own audit reports but does not gather evidence about your workloads.
Question 27Generative AI Fundamentals
A travel agency wants hundreds of tailored email descriptions for different traveler types each week. Which advantage of generative AI does this use MOST?
Correct answer: B. Personalized content at scale
Generative AI can quickly produce many variations of content tailored to different audiences, which saves time over writing each one by hand. It does not guarantee factual accuracy, so outputs still need review.
Generative AI can quickly produce many variations of content tailored to different audiences, which saves time over writing each one by hand. It does not guarantee factual accuracy, so outputs still need review.
Question 28Foundation Model Applications
A team must compare three Bedrock models on accuracy and robustness for a summarization task, using built-in metrics and datasets, without building its own test harness. What should it use?
Correct answer: D. Amazon Bedrock model evaluation
Bedrock model evaluation runs automatic evaluation jobs with built-in metrics and curated or custom datasets, and it can also use human reviewers or a judge model. CloudTrail records API activity and does not measure model quality.
Bedrock model evaluation runs automatic evaluation jobs with built-in metrics and curated or custom datasets, and it can also use human reviewers or a judge model. CloudTrail records API activity and does not measure model quality.
Question 29AI & ML Fundamentals
A team built a model that predicts delivery time in minutes. Which metric is MOST appropriate to evaluate it?
Correct answer: A. Root mean squared error
RMSE measures the typical size of prediction errors for continuous numeric outputs, so it suits a regression model such as delivery time. F1 score, AUC, and precision are classification metrics.
RMSE measures the typical size of prediction errors for continuous numeric outputs, so it suits a regression model such as delivery time. F1 score, AUC, and precision are classification metrics.
Question 30Generative AI Fundamentals
A bank needs to summarize 2 million archived documents with a Bedrock model. Results can arrive within a day, and lower cost matters more than speed. Which option fits BEST?
Correct answer: B. Bedrock batch inference
Bedrock batch inference processes large sets of prompts stored in Amazon S3 asynchronously and is priced lower than on-demand calls for supported models. Real-time calls in a loop work but cost more and need extra code to handle throttling and retries.
Bedrock batch inference processes large sets of prompts stored in Amazon S3 asynchronously and is priced lower than on-demand calls for supported models. Real-time calls in a loop work but cost more and need extra code to handle throttling and retries.
Question 31Foundation Model Applications
A developer wants the model to pick each next token only from the 50 most likely candidates. Which parameter should be set?
Correct answer: A. Top-k
Top-k limits sampling to a fixed number of the most likely tokens, here 50. Top-p instead uses a cumulative probability cutoff, so the number of candidates changes from step to step.
Top-k limits sampling to a fixed number of the most likely tokens, here 50. Top-p instead uses a cumulative probability cutoff, so the number of candidates changes from step to step.
Question 32Responsible AI
A medical information site wants every statement its AI assistant makes to be true and accurate. Which responsible AI dimension does this describe?
Correct answer: D. Veracity
Veracity means outputs are truthful and accurate, which is vital where wrong facts can cause harm. Inclusivity is about serving diverse users well, not about whether statements are correct.
Veracity means outputs are truthful and accurate, which is vital where wrong facts can cause harm. Inclusivity is about serving diverse users well, not about whether statements are correct.
Question 33AI & ML Fundamentals
A fintech team wants one number that shows how well a fraud classifier separates fraud from normal transactions across all possible decision thresholds. Which metric fits?
Correct answer: A. AUC
AUC, the area under the ROC curve, summarizes classifier performance across all thresholds, where 1.0 is perfect separation and 0.5 is random guessing. Accuracy is measured at one threshold and can be misleading on imbalanced fraud data.
AUC, the area under the ROC curve, summarizes classifier performance across all thresholds, where 1.0 is perfect separation and 0.5 is random guessing. Accuracy is measured at one threshold and can be misleading on imbalanced fraud data.
Question 34Security & Governance
A developer puts an admin password in a chatbot's system prompt and tells the model to share it only with managers. What is the MAIN risk?
Correct answer: C. Users could trick the model into revealing it
Anything in a prompt can leak, and a crafted user message could trick the model into revealing the password, so secrets belong in a service such as AWS Secrets Manager, never in prompts. Bedrock does not use prompts to train base models, so that is not the concern.
Anything in a prompt can leak, and a crafted user message could trick the model into revealing the password, so secrets belong in a service such as AWS Secrets Manager, never in prompts. Bedrock does not use prompts to train base models, so that is not the concern.
Question 35Foundation Model Applications
An app asks a model for one FAQ answer, but the model keeps writing extra made-up questions after it. The team wants output to end as soon as the model writes 'Q:'. Which setting fits?
Correct answer: A. A stop sequence
A stop sequence tells the model to stop generating as soon as a set string appears, which ends the response at the right point. A larger context window allows more input and output and does not prevent the extra text.
A stop sequence tells the model to stop generating as soon as a set string appears, which ends the response at the right point. A larger context window allows more input and output and does not prevent the extra text.
Question 36Generative AI Fundamentals
A marketing team wants to create product banner images from text descriptions using an Amazon foundation model in Bedrock. Which model should it choose?
Correct answer: D. Amazon Nova Canvas
Amazon Nova Canvas is an image generation model in Bedrock that creates images from text prompts. Amazon Nova Micro is a text-only model, and Titan Text Embeddings outputs vectors, not images.
Amazon Nova Canvas is an image generation model in Bedrock that creates images from text prompts. Amazon Nova Micro is a text-only model, and Titan Text Embeddings outputs vectors, not images.
Question 37AI & ML Fundamentals
A company stores customer support emails, chat logs, and call audio files. How is this data BEST classified?
Correct answer: B. Unstructured data
Free text, chat logs, and audio do not follow a fixed rows-and-columns format, so they are unstructured data. Structured data fits neatly into tables, such as a spreadsheet of order IDs and amounts.
Free text, chat logs, and audio do not follow a fixed rows-and-columns format, so they are unstructured data. Structured data fits neatly into tables, such as a spreadsheet of order IDs and amounts.
Question 38Foundation Model Applications
A company gets great results from a large Bedrock model but needs lower cost and latency for one specific use case. It wants a smaller model to learn from the large model's responses. Which Bedrock feature fits?
Correct answer: C. Model distillation
Bedrock Model Distillation uses a large teacher model to generate responses that train a smaller student model, giving faster and cheaper inference for a targeted use case. Continued pre-training adds domain knowledge from unlabeled text but does not transfer skill from a larger model.
Bedrock Model Distillation uses a large teacher model to generate responses that train a smaller student model, giving faster and cheaper inference for a targeted use case. Continued pre-training adds domain knowledge from unlabeled text but does not transfer skill from a larger model.
Question 39Generative AI Fundamentals
A product manager with no coding skills wants to quickly compare how different Bedrock models answer the same prompt. What should the manager use?
Correct answer: B. The Bedrock playgrounds in the console
The Bedrock console playgrounds let users enter prompts, pick models, and adjust settings without writing code. A SageMaker training job builds or trains models and requires code, which is far more than the manager needs.
The Bedrock console playgrounds let users enter prompts, pick models, and adjust settings without writing code. A SageMaker training job builds or trains models and requires code, which is far more than the manager needs.
Question 40Responsible AI
A retailer wants to fine-tune a model on customer chat logs that contain names and home addresses. Which risk should it address FIRST?
Correct answer: D. Privacy of personal data in the logs
Training on personal data without proper consent, minimization, or masking can expose that data and break privacy laws, so the retailer should remove or mask personal details first. Latency and settings are secondary concerns.
Training on personal data without proper consent, minimization, or masking can expose that data and break privacy laws, so the retailer should remove or mask personal details first. Latency and settings are secondary concerns.
Question 41Security & Governance
A company policy says raw training data must be deleted 18 months after collection. The data is stored in Amazon S3. What is the simplest way to enforce this?
Correct answer: B. An S3 Lifecycle expiration rule
S3 Lifecycle rules can automatically expire objects after a set age, which enforces the retention policy with no ongoing effort. CloudTrail only logs activity and does not delete data.
S3 Lifecycle rules can automatically expire objects after a set age, which enforces the retention policy with no ongoing effort. CloudTrail only logs activity and does not delete data.
Question 42Foundation Model Applications
A support bot sometimes invents answers when the retrieved documents do not cover a question. Which prompt instruction BEST reduces this?
Correct answer: C. Use only the context, else say you don't know
Telling the model to rely only on supplied context and to admit when it lacks an answer reduces made-up responses. A higher temperature increases randomness and tends to make hallucinations more likely.
Telling the model to rely only on supplied context and to admit when it lacks an answer reduces made-up responses. A higher temperature increases randomness and tends to make hallucinations more likely.
Question 43AI & ML Fundamentals
A social media startup must automatically detect unsafe content, such as violence, in user-uploaded images. Which AWS service fits with the least ML expertise?
Correct answer: A. Amazon Rekognition
Amazon Rekognition includes content moderation APIs that detect unsafe or inappropriate content in images and video. Amazon Textract extracts text and structure from documents and does not judge image content.
Amazon Rekognition includes content moderation APIs that detect unsafe or inappropriate content in images and video. Amazon Textract extracts text and structure from documents and does not judge image content.
Question 44Generative AI Fundamentals
A developer wants to call different Bedrock models with the same request format, so switching models needs only small code changes. Which Bedrock feature helps?
Correct answer: D. The Converse API
The Bedrock Converse API offers one consistent message format across supported models, so code written for one model works with others with few changes. Provisioned Throughput reserves capacity but does not change the request format.
The Bedrock Converse API offers one consistent message format across supported models, so code written for one model works with others with few changes. Provisioned Throughput reserves capacity but does not change the request format.
Question 45Foundation Model Applications
A company updates PDFs in the S3 bucket used as a Bedrock knowledge base data source, but answers still reflect the old content. What should it do?
Correct answer: A. Sync the data source to ingest the changes
A knowledge base reflects new or changed source documents after you run a sync, which is an ingestion job that updates the vector store. Fine-tuning changes model behavior, but it is not how a knowledge base picks up new documents.
A knowledge base reflects new or changed source documents after you run a sync, which is an ingestion job that updates the vector store. Fine-tuning changes model behavior, but it is not how a knowledge base picks up new documents.
Question 46Responsible AI
A voice assistant works well for some English accents but often misunderstands others. Which responsible AI goal should guide the fix?
Correct answer: C. Inclusivity, using diverse speech data
Inclusivity means AI systems serve people from different backgrounds well, which requires training and testing data that covers many accents. Publishing model size adds transparency but does not improve recognition for underserved speakers.
Inclusivity means AI systems serve people from different backgrounds well, which requires training and testing data that covers many accents. Publishing model size adds transparency but does not improve recognition for underserved speakers.
Question 47Generative AI Fundamentals
A company deploys Amazon Q Business over its document repositories with access controls set up correctly. An intern asks about executive pay, which is stored in a restricted folder. What happens?
Correct answer: D. It answers only from content the intern can access
Amazon Q Business can honor the access control lists from connected data sources, so users only get answers from documents they are allowed to see. It does not retrain a model per user; it filters which content is used for each response.
Amazon Q Business can honor the access control lists from connected data sources, so users only get answers from documents they are allowed to see. It does not retrain a model per user; it filters which content is used for each response.
Question 48Security & Governance
A team wants automatic checks for missing values and out-of-range fields in the data pipelines that feed its ML models. Which AWS service fits BEST?
Correct answer: D. AWS Glue Data Quality
AWS Glue Data Quality lets you define rules that check data for issues such as missing values or invalid ranges and reports quality scores. Catching these problems early protects model accuracy, which AWS Shield, a DDoS protection service, cannot do.
AWS Glue Data Quality lets you define rules that check data for issues such as missing values or invalid ranges and reports quality scores. Catching these problems early protects model accuracy, which AWS Shield, a DDoS protection service, cannot do.
Question 49AI & ML Fundamentals
A hotel group wants to know whether thousands of guest reviews are positive, negative, neutral, or mixed. Which AWS service should it use?
Correct answer: B. Amazon Comprehend
Amazon Comprehend is a natural language processing service that detects sentiment, entities, and key phrases in text. Amazon Translate converts text between languages but does not judge sentiment.
Amazon Comprehend is a natural language processing service that detects sentiment, entities, and key phrases in text. Amazon Translate converts text between languages but does not judge sentiment.
Question 50Foundation Model Applications
An airline launched a generative AI assistant to reduce calls to human agents. Which measure BEST shows whether it delivers business value?
Correct answer: C. The share of chats resolved with no agent
A business metric such as the self-service resolution rate ties the assistant directly to the goal of fewer agent calls. ROUGE measures text overlap with references and does not show whether customers got their issue solved.
A business metric such as the self-service resolution rate ties the assistant directly to the goal of fewer agent calls. ROUGE measures text overlap with references and does not show whether customers got their issue solved.
Question 51Generative AI Fundamentals
A company already fine-tuned an open-weight Llama model elsewhere and wants to run it serverlessly next to other Bedrock models through the same API. Which feature fits?
Correct answer: D. Bedrock Custom Model Import
Bedrock Custom Model Import lets you bring supported custom model weights, such as a fine-tuned Llama model, into Bedrock and invoke them through the same API. Knowledge Bases connect models to your documents for retrieval and do not host your own model weights.
Bedrock Custom Model Import lets you bring supported custom model weights, such as a fine-tuned Llama model, into Bedrock and invoke them through the same API. Knowledge Bases connect models to your documents for retrieval and do not host your own model weights.
Question 52Foundation Model Applications
A legal assistant app sends the same 20-page policy document at the start of every prompt, followed by a short question. Which Bedrock feature can cut cost and latency for the repeated part?
Correct answer: C. Prompt caching
Bedrock prompt caching stores a repeated prompt prefix so later requests can reuse it, which lowers input token cost and speeds up responses on supported models. Provisioned Throughput reserves capacity but does not avoid reprocessing the same long prefix.
Bedrock prompt caching stores a repeated prompt prefix so later requests can reuse it, which lowers input token cost and speeds up responses on supported models. Provisioned Throughput reserves capacity but does not avoid reprocessing the same long prefix.
Question 53AI & ML Fundamentals
Two teams compute the same customer features in different ways, so models behave differently in training and production. Which SageMaker AI capability helps solve this?
Correct answer: B. SageMaker Feature Store
SageMaker Feature Store is a central repository to store, share, and reuse features, so training and inference use the same definitions. Data Wrangler helps prepare data, but it is not the shared source of features across teams.
SageMaker Feature Store is a central repository to store, share, and reuse features, so training and inference use the same definitions. Data Wrangler helps prepare data, but it is not the shared source of features across teams.
Question 54Responsible AI
A team wants to understand the intended uses, limitations, and responsible AI design choices of an AWS managed AI service it plans to adopt. Which AWS resource provides this?
Correct answer: C. AWS AI Service Cards
AWS AI Service Cards describe intended use cases, limitations, and responsible AI design choices for AWS AI services. AWS Artifact SOC reports cover security controls, not how an AI service should be used responsibly.
AWS AI Service Cards describe intended use cases, limitations, and responsible AI design choices for AWS AI services. AWS Artifact SOC reports cover security controls, not how an AI service should be used responsibly.
Question 55Security & Governance
Before using an S3 data lake to fine-tune a model, a company wants to find files that contain personal data such as credit card numbers. Which service fits BEST?
Correct answer: A. Amazon Macie
Amazon Macie uses machine learning and pattern matching to discover sensitive data, such as PII and financial information, in Amazon S3. CloudTrail logs API activity but does not inspect file contents.
Amazon Macie uses machine learning and pattern matching to discover sensitive data, such as PII and financial information, in Amazon S3. CloudTrail logs API activity but does not inspect file contents.
Question 56Generative AI Fundamentals
A product page says a model has 70 billion parameters. What are parameters?
Correct answer: B. The learned weights inside the model
Parameters are the internal weights a model adjusts during training, and the parameter count is a rough measure of model size. More parameters often means more capability but also higher cost and latency, while prompt length limits come from the context window.
Parameters are the internal weights a model adjusts during training, and the parameter count is a rough measure of model size. More parameters often means more capability but also higher cost and latency, while prompt length limits come from the context window.
Question 57Foundation Model Applications
A team wants to link a prompt, a knowledge base lookup, and a Lambda function into one visual, reusable generative AI workflow. Which Bedrock feature fits?
Correct answer: A. Amazon Bedrock Flows
Amazon Bedrock Flows lets you connect prompts, knowledge bases, Lambda functions, and other steps into a workflow with a visual builder and run it as one unit. Batch inference processes many prompts offline but does not orchestrate multi-step logic.
Amazon Bedrock Flows lets you connect prompts, knowledge bases, Lambda functions, and other steps into a workflow with a visual builder and run it as one unit. Batch inference processes many prompts offline but does not orchestrate multi-step logic.
Question 58AI & ML Fundamentals
A marketing analyst with no coding skills wants to build a churn prediction model from a spreadsheet using a point-and-click interface. Which option fits BEST?
Correct answer: C. SageMaker Canvas
SageMaker Canvas lets business users build ML models and generate predictions through a no-code visual interface. Training jobs written in Python require coding skills the analyst does not have.
SageMaker Canvas lets business users build ML models and generate predictions through a no-code visual interface. Training jobs written in Python require coding skills the analyst does not have.
Question 59Generative AI Fundamentals
During traffic peaks, a Bedrock app gets throttled in its home Region. The team wants Bedrock to route requests to other Regions automatically for more throughput. Which feature fits?
Correct answer: D. Cross-Region inference
Cross-Region inference uses inference profiles to spread requests across several AWS Regions, which helps absorb traffic bursts without throttling. Model invocation logging records requests and responses but adds no capacity.
Cross-Region inference uses inference profiles to spread requests across several AWS Regions, which helps absorb traffic bursts without throttling. Model invocation logging records requests and responses but adds no capacity.
Question 60Foundation Model Applications
In a retrieval augmented generation app, what happens to the user's question before the vector store is searched?
Correct answer: D. It is turned into a vector by the embedding model
The question is embedded with the same model used for the documents, so the vector store can find passages whose vectors are closest in meaning. Fine-tuning happens offline and never runs on each user question.
The question is embedded with the same model used for the documents, so the vector store can find passages whose vectors are closest in meaning. Fine-tuning happens offline and never runs on each user question.
Question 61Responsible AI
A lender's model was fair at launch. As the applicant population changes, the lender wants to detect whether bias grows over time in production. Which approach fits BEST?
Correct answer: A. Bias drift monitoring in Model Monitor
SageMaker Model Monitor integrates with Clarify to track bias metrics on live data and alert when they drift past thresholds. A one-time check before training cannot catch bias that appears later as real-world data changes.
SageMaker Model Monitor integrates with Clarify to track bias metrics on live data and alert when they drift past thresholds. A one-time check before training cannot catch bias that appears later as real-world data changes.
Question 62Security & Governance
A company wants to show customers that its AI practices follow an international standard written specifically for AI management systems. Which standard is this?
Correct answer: B. ISO/IEC 42001
ISO/IEC 42001 is the international standard for AI management systems, covering governance and the responsible development and use of AI. ISO/IEC 27001 covers information security management in general, not AI specifically.
ISO/IEC 42001 is the international standard for AI management systems, covering governance and the responsible development and use of AI. ISO/IEC 27001 covers information security management in general, not AI specifically.
Question 63AI & ML Fundamentals
A pizza chain wants customers to place orders through a chatbot by text or voice. Which AWS service is built for this conversational interface?
Correct answer: D. Amazon Lex
Amazon Lex builds conversational interfaces with speech and text, handling intents and collecting the details needed to finish a task such as an order. Amazon Polly only converts text to speech and cannot manage a conversation.
Amazon Lex builds conversational interfaces with speech and text, handling intents and collecting the details needed to finish a task such as an order. Amazon Polly only converts text to speech and cannot manage a conversation.
Question 64Generative AI Fundamentals
Model providers often improve a pretrained LLM by having people rank its answers and then training it to prefer the higher-ranked ones. What is this technique called?
Correct answer: C. Reinforcement learning from human feedback
Reinforcement learning from human feedback, or RLHF, uses human preference rankings to build a reward signal that steers the model toward helpful and safe answers. Continued pre-training uses unlabeled domain text and does not rely on human rankings.
Reinforcement learning from human feedback, or RLHF, uses human preference rankings to build a reward signal that steers the model toward helpful and safe answers. Continued pre-training uses unlabeled domain text and does not rely on human rankings.
Question 65Foundation Model Applications
A team is choosing between RAG and fine-tuning for a help desk assistant. What is a real trade-off of RAG?
Correct answer: B. Prompts grow with retrieved text, adding token cost
RAG adds retrieved passages to each prompt, so requests use more input tokens and take slightly longer, although no training is needed. Retraining model weights is a cost of fine-tuning, not of RAG.
RAG adds retrieved passages to each prompt, so requests use more input tokens and take slightly longer, although no training is needed. Retraining model weights is a cost of fine-tuning, not of RAG.
The real AWS AI Practitioner exam at a glance
65questions
1h 30mtime limit
MCQmultiple choice
80%+aim for here
Score 80% or more on two different sets before you book the real exam. Need a plan for your weak topics? See how we help you pass, or get a full refund.
More free AWS practice tests
AWS Cloud Practitioner Practice Test65 questions · CLF-C02AWS Solutions Architect Associate Practice Test65 questions · SAA-C03AWS Solutions Architect Professional Practice Test75 questions · SAP-C03AWS Security Specialty Practice Test65 questions · SCS-C03AWS DevOps Engineer Professional Practice Test75 questions · DOP-C02AWS Data Engineer Associate Practice Test65 questions · DEA-C01
See all 36 free practice tests →
Original practice questions written by CertOcean in the style of the AWS AI Practitioner exam. They are not real exam questions. AWS AI Practitioner and Amazon Web Services are trademarks of their owners. Real exam format: 65 questions (50 scored, 15 unscored) · 90 minutes · passing score 700 on a 100-1000 scale.