CertSafari

    Free AWS Certified AI Practitioner (AIF-C01) Sample Questions

    35 free sample questions from our bank of 397+, covering every exam domain, with answers and detailed explanations. Updated November 2026.

    Domain 1: Fundamentals of AI and ML

    Subdomain 1.1: Explain basic AI concepts and terminologies.

    1.A company aggregates two datasets: a spreadsheet with defined columns for customer ID, purchase date, and amount, and a folder of free-form written product review text submitted by customers with no fixed structure. Which two data types are being described, respectively? (Select TWO.)(Select 2)

    1. A.Tabular data
    2. B.Unstructured text data
    3. C.Time-series data
    4. D.Labeled data
    5. E.Image data
    Show answer & explanation

    Correct answers: A, B — Tabular data; Unstructured text data

    • A. This is correct because tabular data is organized into rows and defined columns, matching the spreadsheet with customer ID, purchase date, and amount fields.
    • B. This is correct because unstructured text data has no fixed fields or schema, matching the free-form written product review text submitted by customers.
    • C. This is incorrect for this scenario because time-series data is defined by its ordering along a time dimension as the primary organizing factor, which is not how either dataset is described here.
    • D. This is incorrect for this scenario because labeled data specifically refers to examples paired with a known target or output value, and neither dataset is described as having such labels attached.
    • E. This is incorrect for this scenario because image data consists of pixel-based visual files, and neither dataset described here contains images.

    1.3: Describe the AI/ML development lifecycle.

    2.A company's leadership is reviewing whether a newly deployed ML model justifies its ongoing investment. Which of the following are examples of business metrics, as opposed to model performance metrics, that leadership should examine? (Select TWO.)(Select 2)

    1. A.Return on investment
    2. B.Cost per user
    3. C.Area Under the Curve
    4. D.F1 score
    5. E.Accuracy
    Show answer & explanation

    Correct answers: A, B — Return on investment; Cost per user

    • A. Return on investment is correct because it is a business metric that measures the financial value the model generates relative to what was spent to build and run it.
    • B. Cost per user is correct because it is a business metric reflecting the operational expense of serving the model to each user, rather than a statistical measure of prediction quality.
    • C. Area Under the Curve is incorrect because it is a model performance metric describing a classifier's ability to distinguish between classes, not a financial or business measure.
    • D. F1 score is incorrect because it is a model performance metric balancing precision and recall, not a business-level measure of value or cost.
    • E. Accuracy is incorrect because it is a model performance metric describing the proportion of correct predictions, not a business outcome measure.

    Subdomain 1.1: Explain basic AI concepts and terminologies.

    3.A manufacturer installs cameras above an assembly line and wants a system that scans each captured photo of a circuit board to automatically flag scratches, missing components, or misalignment before shipping. Which AI capability should the manufacturer use?

    1. A.Computer vision that inspects the pixel content of each board image and marks damaged, absent or shifted parts
    2. B.Natural language processing that classifies the written shift notes technicians attach to each finished board
    3. C.Time-series forecasting that projects the coming quarter's defect rate from past line output and rejection counts
    4. D.Speech recognition that transcribes the audio captured on the factory floor into searchable text for supervisors
    Show answer & explanation

    Correct answer: A — Computer vision that inspects the pixel content of each board image and marks damaged, absent or shifted parts

    • A. Computer vision is the AI capability for interpreting image data. It can examine each photo of a circuit board and detect surface scratches, missing components and parts that sit out of position.
    • B. Natural language processing works on written or spoken language. The manufacturer needs each captured photo of a board scanned automatically, and a photograph is not text, so a language-based capability cannot analyze it. Classifying notes that people write about a product does not inspect the board image itself.
    • C. Time-series forecasting predicts aggregate figures from historical data. The manufacturer needs a decision on each board from its own photo, and a projected quarterly defect rate cannot show whether one board has a scratch or a missing part.
    • D. Speech recognition converts spoken audio into text. The cameras above the line produce images, not audio, so transcripts of floor recordings contain nothing about the condition of any board.

    1.3: Describe the AI/ML development lifecycle.

    4.A data science team is testing several algorithms and hyperparameter combinations before committing to a production model, tracking each run's metrics for comparison. Which MLOps concept does this describe?

    1. A.Experimentation, the practice of training and evaluating model variants to pick a candidate for release
    2. B.Model monitoring, the practice of tracking a deployed model's prediction quality and drift over time
    3. C.Model versioning, the practice of registering each trained artifact with its lineage for rollback
    4. D.Production readiness, the practice of validating a model against operational requirements before launch
    Show answer & explanation

    Correct answer: A — Experimentation, the practice of training and evaluating model variants to pick a candidate for release

    • A. Training and evaluating many algorithm and hyperparameter variants, and recording each run's results so they can be compared, is experimentation. It happens before a model is selected for production.
    • B. Model monitoring watches a model that is already deployed and serving predictions. The team in this scenario has no live model yet and is still comparing candidates to decide which one to ship.
    • C. Model versioning catalogs trained artifacts with their lineage so they can be reproduced or rolled back. It does not involve trying out different algorithms and hyperparameter settings to see which performs best.
    • D. Production readiness validation applies once a model has already been chosen, to confirm it meets operational requirements. This team is still choosing among candidates, so that step has not been reached.

    Subdomain 1.1: Explain basic AI concepts and terminologies.

    5.An airline generates personalized seat upgrade offers for customers browsing its website and must return a prediction within milliseconds of the request. Which type of inferencing best fits this requirement?

    1. A.Real-time inferencing, scoring each request on a hosted endpoint and sending the prediction straight back to the caller
    2. B.Batch inferencing, scoring collected records together as one job and writing the predictions to storage at the end
    3. C.Asynchronous inferencing, placing each request in a queue and returning the prediction once processing completes
    4. D.Offline inferencing, scoring stored profiles on a nightly schedule and giving staff the results to review
    Show answer & explanation

    Correct answer: A — Real-time inferencing, scoring each request on a hosted endpoint and sending the prediction straight back to the caller

    • A. Real-time inferencing keeps a model deployed behind a live endpoint that scores individual requests as they arrive and replies within milliseconds. That matches an offer that has to be produced while the customer is on the page.
    • B. Batch inferencing is a real mode, but it processes a collected set of records in a single job and writes the output when the job finishes. It cannot return a prediction within milliseconds of one customer's request.
    • C. Asynchronous inferencing is a real mode, built for large payloads or long processing times. Requests wait in a queue and the result is fetched later, so it gives no millisecond response to a customer who is browsing.
    • D. Offline inferencing scores stored data on a schedule and produces results for later review. The delay between a customer's request and the prediction is far longer than the millisecond requirement allows.

    Subdomain 1.2: Identify practical use cases for AI.

    6.Which AWS AI service uses computer vision to detect objects, scenes, and unsafe content within images and videos?

    1. A.Amazon Textract
    2. B.Amazon Rekognition
    3. C.Amazon Comprehend
    4. D.Amazon Forecast
    Show answer & explanation

    Correct answer: B — Amazon Rekognition

    • A. Amazon Textract extracts printed and handwritten text, forms, and tables from scanned documents; it does not analyze image content such as objects or scenes.
    • B. Amazon Rekognition is correct because it applies computer vision to identify objects, scenes, activities, and inappropriate content in images and video.
    • C. Amazon Comprehend performs natural language processing on text, such as sentiment and entity extraction, rather than analyzing image or video content.
    • D. Amazon Forecast generates time-series predictions, such as demand or inventory forecasts, from historical data. It has no capability for analyzing image or video content such as objects, scenes, or unsafe material.

    1.3: Describe the AI/ML development lifecycle.

    7.A company evaluating a binary classifier for loan default prediction wants a single metric that balances both precision and recall, since the dataset has far more non-default cases than default cases. Which metric should the team prioritize?

    1. A.F1 score
    2. B.Overall accuracy
    3. C.Recall score
    4. D.True negative rate
    Show answer & explanation

    Correct answer: A — F1 score

    • A. F1 score is correct because it is the harmonic mean of precision and recall, making it a more informative single metric than accuracy when classes are imbalanced.
    • B. Overall accuracy is incorrect because with a highly imbalanced dataset, a model can achieve high accuracy simply by predicting the majority class and still miss most true defaults.
    • C. Recall score is incorrect because it measures only the share of actual defaults that are caught. It ignores how many of the flagged loans are false alarms, so it does not balance precision and recall, and a model that flags every loan would score perfectly.
    • D. True negative rate is incorrect because it measures how well the model identifies non-defaults, which is the majority class. It uses neither precision nor recall, so it cannot balance them and says little about how well defaults are found.

    1.3: Describe the AI/ML development lifecycle.

    8.Which of the following are core concepts of ML operations (MLOps)? (Select TWO.)(Select 2)

    1. A.Automating training and deployment steps as repeatable pipelines.
    2. B.Deploying each approved model by hand, following a documented checklist.
    3. C.Watching production model metrics and retraining when quality declines.
    4. D.Letting each data scientist keep model artifacts in their own workspace.
    5. E.Freezing the deployed model after launch to keep predictions consistent.
    Show answer & explanation

    Correct answers: A, C — Automating training and deployment steps as repeatable pipelines.; Watching production model metrics and retraining when quality declines.

    • A. Correct. Turning training and deployment into repeatable, automated pipelines is a foundational MLOps concept. It cuts manual effort and human error, and it makes every model release consistent and reproducible.
    • B. Incorrect. A checklist makes manual deployment more orderly, but each release is still a hand-run, error-prone process. MLOps replaces that with automated, repeatable deployment pipelines.
    • C. Correct. Tracking how a production model performs and retraining it when quality declines is a core MLOps practice. It addresses issues such as data drift and keeps model quality high over time.
    • D. Incorrect. Artifacts held in individual workspaces have no shared versioning, lineage or access control, so results are hard to reproduce or audit. MLOps centralizes them in managed stores such as a model registry.
    • E. Incorrect. Consistent predictions do not mean correct ones: live data shifts, so a frozen model gradually loses accuracy. MLOps monitors deployed models and retrains them instead of leaving them unchanged.

    Subdomain 1.2: Identify practical use cases for AI.

    9.A company wants to draft first-pass replies to long, varied customer emails. It has no labeled training data and only a small ML team. Which approach best fits this use case?

    1. A.Prompt a Bedrock foundation model with each email and a reply-writing instruction, then return its output
    2. B.Train a regression model in SageMaker AI on past email lengths, then return its predicted reply text
    3. C.Cluster emails by topic in SageMaker AI, then return the stored template for the matching cluster
    4. D.Train a classifier in SageMaker AI to tag emails by intent, then return the canned reply for that intent
    Show answer & explanation

    Correct answer: A — Prompt a Bedrock foundation model with each email and a reply-writing instruction, then return its output

    • A. A pretrained foundation model in Amazon Bedrock can write a tailored reply to each long, varied email from a prompt alone. It needs no labeled data and no training pipeline, so a small ML team can run it.
    • B. Regression predicts continuous numeric values, such as a length or a score. Trained on past email lengths, it can output a number but cannot write the wording of a reply to a long, varied email.
    • C. Clustering only groups similar emails. It writes no text, so each customer gets the stored template for their cluster, not a reply that addresses what their long, varied email says.
    • D. A supervised classifier needs labeled examples of each intent, and the company has no labeled data. Even a well-trained one returns a fixed reply per intent, which cannot address the specifics of long, varied emails.

    Domain 2: Fundamentals of GenAI

    Subdomain 2.3: Describe AWS infrastructure and technologies for building GenAI applications.

    10.A retail company wants to quickly add a natural-language customer support chatbot to their website by calling a foundation model through a simple API, without provisioning any servers or managing GPU clusters. Which AWS service should they use?

    1. A.Amazon SageMaker AI
    2. B.Amazon Bedrock
    3. C.AWS Lambda
    4. D.Amazon EC2
    Show answer & explanation

    Correct answer: B — Amazon Bedrock

    • A. SageMaker AI centers on building, training, and hosting custom machine learning models with more infrastructure configuration, which is more than this team needs for calling a pretrained foundation model.
    • B. This is correct because it is a fully managed, serverless service that provides API access to foundation models without any server or cluster management.
    • C. This is a general-purpose serverless compute service for running code in response to events; it is not itself a source of foundation models.
    • D. This requires provisioning and managing virtual servers, which does not match the serverless, no-infrastructure requirement in the scenario.

    Subdomain 2.3: Describe AWS infrastructure and technologies for building GenAI applications.

    11.What best describes Strands Agents?

    1. A.An open-source SDK that lets developers build AI agents using a model-driven approach with only a few lines of code.
    2. B.A fully managed relational database engine used to store transactional agent conversation logs at scale.
    3. C.A proprietary GUI-only tool that requires no code but only supports a single hardcoded foundation model.
    4. D.A network load balancer service that distributes incoming HTTP requests across a fleet of EC2 instances.
    Show answer & explanation

    Correct answer: A — An open-source SDK that lets developers build AI agents using a model-driven approach with only a few lines of code.

    • A. This is correct because it is an open-source SDK that simplifies building AI agents through a model-driven approach with minimal code.
    • B. This describes a managed database engine for storing transactional data, which is not what this lightweight agent-building SDK provides.
    • C. This describes a no-code, single-model tool, which contradicts the SDK's code-based, model-flexible design.
    • D. This describes a load balancing service for distributing network traffic, which is unrelated to building AI agents.

    2.1: Explain the basic concepts of generative AI (GenAI).

    12.What best describes the token-based pricing model commonly used for foundation model inference?

    1. A.Charges are based on hours of reserved model capacity, billed whether or not requests arrive
    2. B.Charges are based on the input and output tokens processed, metered for each request
    3. C.Charges are based on the number of API calls, at one rate for short and long prompts
    4. D.Charges are based on the output tokens generated, with prompt tokens left unbilled
    Show answer & explanation

    Correct answer: B — Charges are based on the input and output tokens processed, metered for each request

    • A. Paying for reserved capacity by the hour is a provisioned-throughput style of billing. Cost follows the time capacity is held, not the tokens processed, so it is not token-based pricing.
    • B. Token-based pricing meters the input tokens sent and the output tokens generated in each request. Longer prompts and longer responses therefore cost more.
    • C. A per-call rate ignores how much text each request carries, so a short prompt and a long one cost the same. Token-based pricing instead scales with the tokens processed.
    • D. Token-based pricing bills the input tokens as well as the output tokens. Leaving prompt tokens unbilled misdescribes the model, because long prompts add to the cost of a request.

    Subdomain 2.2: Understand the capabilities and limitations of GenAI for solving business problems.

    13.A product team deployed a generative AI writing assistant and wants to measure whether it is actually worth the ongoing subscription cost. Which metric would most directly answer that business question?

    1. A.Return on investment (ROI)
    2. B.Monthly active users (MAU)
    3. C.Average response time (ms)
    4. D.Net promoter score (NPS)
    Show answer & explanation

    Correct answer: A — Return on investment (ROI)

    • A. Return on investment is correct because it directly compares the value or savings generated by the tool against its cost, answering whether the subscription is worthwhile.
    • B. Monthly active users shows how many people use the assistant, but usage says nothing about the value it produces compared with the subscription fee. A widely used tool can still cost more than it returns.
    • C. Average response time measures how fast the assistant replies, which is a performance property. A fast tool can still cost more than the value it delivers, so this metric cannot show whether the subscription is worth paying for.
    • D. Net promoter score captures how willing users are to recommend the assistant, which is a sentiment measure. It contains no cost or financial benefit data, so it cannot show whether the subscription pays for itself.

    2.1: Explain the basic concepts of generative AI (GenAI).

    14.A team building a customer-support chatbot is loading hundreds of long PDF manuals into a knowledge base. Before creating embeddings, they split each manual into smaller passages of a few paragraphs each so that relevant sections can be retrieved independently. What is this preprocessing step called?

    1. A.Chunking
    2. B.Reranking
    3. C.Tokenizing
    4. D.Embedding
    Show answer & explanation

    Correct answer: A — Chunking

    • A. Splitting long documents into smaller, retrievable passages before embedding them is called chunking, and it allows a retrieval system to return only the most relevant section instead of an entire manual.
    • B. Reranking reorders passages that a search has already retrieved so the most relevant ones come first. It happens at query time, after retrieval, and does not split manuals into passages before embeddings are created.
    • C. Tokenizing converts text into the small units (tokens) a model reads internally, such as words or word fragments. It is far more granular than splitting a manual into passages of a few paragraphs that can be retrieved independently.
    • D. Embedding converts each passage into a numeric vector that captures its meaning. It takes the already-split passages as input and is the step that follows the split, so it does not divide the manuals into passages.

    2.1: Explain the basic concepts of generative AI (GenAI).

    15.A product team is documenting which tasks a generative AI foundation model can typically perform out of the box. Which of the following are common generative AI use cases? (Select TWO.)(Select 2)

    1. A.Creating marketing images from text prompts
    2. B.Summarizing long reports into brief overviews
    3. C.Enforcing firewall rules on network traffic
    4. D.Compiling application source code into binaries
    5. E.Tuning the storage layout of database indexes
    Show answer & explanation

    Correct answers: A, B — Creating marketing images from text prompts; Summarizing long reports into brief overviews

    • A. Producing new images from a written description is a well-known generative AI capability that foundation models handle out of the box. It is widely used for marketing and creative content.
    • B. Condensing long documents into shorter overviews is a core generative AI use case, since foundation models can read lengthy text and produce a concise summary of the key points.
    • C. Enforcing firewall rules is a network-security task. Dedicated firewalls and security services do it through deterministic configuration, so it is not a content-generation capability of a foundation model.
    • D. Turning source code into executable binaries is done by a compiler toolchain, which follows deterministic rules. A foundation model can help write code, but it does not perform compilation, so this is not a generative AI use case.
    • E. Tuning how a database physically stores its indexes is a database-administration task handled by the database engine and its administrators. Foundation models are not built to perform it, so it is not a generative AI use case.

    Subdomain 2.2: Understand the capabilities and limitations of GenAI for solving business problems.

    16.A legal team asks a generative AI assistant to summarize case law and cites a court ruling that does not exist. Which limitation of generative AI does this describe?

    1. A.Hallucination
    2. B.Latency
    3. C.Token pricing
    4. D.Context window
    Show answer & explanation

    Correct answer: A — Hallucination

    • A. A hallucination occurs when a generative model produces confident but fabricated or factually incorrect content, such as citing a court ruling that never existed, and it is a well-known limitation of GenAI.
    • B. Latency refers to the response time of a model and has no bearing on whether the content it generates is factually accurate.
    • C. Token pricing describes how usage is billed based on input and output tokens and does not explain fabricated content in a response.
    • D. A context window is the maximum amount of text a model can take in and work with at once. It limits how much material fits in a prompt. It does not make a model invent a court ruling that was never issued.

    Subdomain 2.3: Describe AWS infrastructure and technologies for building GenAI applications.

    17.A data science team wants to browse a catalog of pretrained, open-source models, incrementally fine-tune a selected model, and deploy it directly from within SageMaker Studio. Which capability should the team use?

    1. A.Amazon SageMaker Autopilot
    2. B.Amazon SageMaker JumpStart
    3. C.Amazon Bedrock Agents
    4. D.Amazon SageMaker Clarify
    Show answer & explanation

    Correct answer: B — Amazon SageMaker JumpStart

    • A. SageMaker Autopilot is an AutoML capability that builds, trains and tunes models from the team's own tabular data. It does not offer a browsable catalog of pretrained open-source models to fine-tune and deploy from SageMaker Studio.
    • B. SageMaker JumpStart is correct because it provides pretrained, open-source models directly inside SageMaker Studio that can be incrementally trained, tuned, and deployed.
    • C. Amazon Bedrock Agents orchestrates multistep tasks by letting a foundation model call APIs and knowledge bases. It is not a catalog of open-source models that a team can fine-tune and deploy from SageMaker Studio.
    • D. SageMaker Clarify detects bias in data and models and explains model predictions. It does not provide a catalog of pretrained open-source models or a way to fine-tune and deploy them from SageMaker Studio.

    2.1: Explain the basic concepts of generative AI (GenAI).

    18.A team is designing an agentic AI application that can call external APIs, retain relevant details across a multi-turn conversation, and decide the order in which to perform sub-tasks. Which of the following are core capabilities this design relies on? (Select all that apply.)(Select 2)

    1. A.Fine-tuning on past chat logs, letting the agent keep users' details between turns
    2. B.Parametric knowledge, letting the agent answer API-related requests from its training data
    3. C.Tool integration, letting the agent invoke external APIs while it works through a task
    4. D.Predefined workflow graphs, letting the agent follow a sub-task order set ahead of time
    5. E.Conversation memory, letting the agent carry earlier details into later turns
    Show answer & explanation

    Correct answers: C, E — Tool integration, letting the agent invoke external APIs while it works through a task; Conversation memory, letting the agent carry earlier details into later turns

    • A. Incorrect. Fine-tuning updates the model's weights in a separate training run, so it cannot pick up details from a conversation that is still in progress. Retaining relevant details across turns needs a memory mechanism that the running agent reads and writes.
    • B. Incorrect. Parametric knowledge is fixed when training ends and can only produce text about APIs. The application has to call external APIs, which requires a tool-calling capability, so training data cannot stand in for it.
    • C. Correct. Tool integration lets the agent invoke external APIs and services during a task, so it can take actions instead of only generating text. The application relies on this capability to call external APIs.
    • D. Incorrect. A predefined workflow graph sets the order of steps at design time, before any request arrives. This design needs the agent to decide the order of sub-tasks itself, which calls for dynamic planning and reasoning rather than a preset path.
    • E. Correct. Conversation memory lets the agent keep earlier details and use them in later turns. The application relies on this capability to retain relevant details across a multi-turn conversation.

    Domain 3: Applications of Foundation Models

    Subdomain 3.1: Describe design considerations for applications that use foundation models (FMs).

    19.A development team is building a chatbot application on Amazon Bedrock that repeatedly sends the same long system prompt and reference document with every API call. Which foundation model selection criterion primarily addresses reducing the cost and latency of processing this repeated content?

    1. A.Model size
    2. B.Prompt caching
    3. C.Multi-lingual support
    4. D.Output length limit
    Show answer & explanation

    Correct answer: B — Prompt caching

    • A. Model size refers to a model's parameter count and general capacity, not to how repeated prompt content is reprocessed on each call.
    • B. Prompt caching lets a model reuse the processed representation of an unchanged prompt prefix, so repeated system prompts and reference documents are not reprocessed on every call, cutting cost and latency.
    • C. Multi-lingual support describes how many languages a model handles well and has no bearing on reprocessing repeated content.
    • D. An output length limit caps how much text is generated in a response and does not affect the cost of reprocessing repeated input content.

    Subdomain 3.1: Describe design considerations for applications that use foundation models (FMs).

    20.An engineer increases the temperature inference parameter before generating marketing taglines with a foundation model. What effect does this change have on the model's output?

    1. A.Responses differ more from run to run
    2. B.Responses are cut off earlier
    3. C.Responses favour the most likely words
    4. D.Responses are more factually accurate
    Show answer & explanation

    Correct answer: A — Responses differ more from run to run

    • A. A higher temperature flattens the probability distribution used for token selection, so less likely tokens get picked more often and the same prompt produces more varied outputs. That suits generating diverse marketing taglines.
    • B. Output length is limited by the maximum token setting and by when the model emits an end-of-sequence token. Temperature only changes how randomly tokens are sampled, so raising it does not make responses end sooner.
    • C. Favouring the highest-probability tokens is what a lower temperature does. Raising the temperature spreads probability toward less likely tokens, so output becomes less predictable, not more.
    • D. Temperature does not add knowledge or improve correctness. Higher values make sampling more random, which tends to raise the chance of loosely grounded wording, so accuracy does not improve.

    Subdomain 3.4: Describe methods to evaluate FM performance.

    21.A team built a retrieval-augmented generation (RAG) application that answers questions using a company knowledge base. During evaluation, they find the generated answers are fluent but frequently unsupported by the retrieved documents. Which aspect of the RAG pipeline should the team focus on evaluating further?

    1. A.Coverage of the facts each answer relies on within the passages the retriever returns
    2. B.Token generation speed of the foundation model relative to the application's response target
    3. C.Query throughput of the vector database relative to the request volume expected at peak
    4. D.Formatting consistency of the retrieved context in the prompt template across user requests
    Show answer & explanation

    Correct answer: A — Coverage of the facts each answer relies on within the passages the retriever returns

    • A. Fluent answers that the retrieved documents do not back up suggest the retrieved passages are missing facts the answers need, so the model fills the gaps on its own. Evaluating how well the retrieved passages cover those facts targets the failure directly.
    • B. Generation speed is a performance property. The reported problem is the content of the answers, and a model that meets its response target can still produce text the retrieved documents do not support.
    • C. Query throughput at peak load shows whether the vector database can handle traffic. It does not show whether the passages it returns contain the facts the answers need, which is the reported symptom.
    • D. Consistent formatting of the context in the prompt affects how it is presented to the model. It does not show whether the retrieved content itself supports the claims in the generated answers.

    Subdomain 3.3: Describe the training and fine-tuning process for FMs.

    22.A retailer fine-tunes a customer service model almost entirely on English-language chat transcripts from one region, then deploys it globally and finds it performs poorly on non-English and region-specific queries. Which data preparation issue most directly caused this outcome?

    1. A.Sparse labeling of intents and outcomes across the fine-tuning transcripts
    2. B.Narrow language and regional coverage across the fine-tuning transcripts
    3. C.Strict data governance retention limits applied to the fine-tuning transcripts
    4. D.Heavy reliance on reinforcement learning from human feedback during fine-tuning
    Show answer & explanation

    Correct answer: B — Narrow language and regional coverage across the fine-tuning transcripts

    • A. Sparse labeling would weaken the model's answers regardless of language or region. The reported pattern is poor results specifically on non-English and region-specific queries. That points to which languages and regions the transcripts covered, not to how well they were labeled.
    • B. Transcripts drawn almost entirely from one region and language leave the training data unrepresentative of the customers served worldwide. The model saw little non-English text or regional phrasing, so it performs poorly on those queries once deployed globally.
    • C. Retention limits could shrink how much transcript history is available, but they do not decide which languages or regions that history comes from. Even with more data retained, it would still be English and from one region, so this does not explain the gap.
    • D. Reinforcement learning from human feedback is a tuning method, not a data preparation issue. Relying on it heavily would not narrow the languages and regions the transcripts cover, so it cannot account for failures on non-English and regional queries.

    Subdomain 3.1: Describe design considerations for applications that use foundation models (FMs).

    23.Which AWS services can be used to store embeddings as the vector store for a Retrieval Augmented Generation solution built with Amazon Bedrock Knowledge Bases? (Select THREE.)(Select 3)

    1. A.Amazon OpenSearch Service
    2. B.Amazon Aurora
    3. C.Amazon Neptune
    4. D.AWS Glue
    5. E.Amazon DynamoDB
    6. F.Amazon Elastic File System
    Show answer & explanation

    Correct answers: A, B, C — Amazon OpenSearch Service; Amazon Aurora; Amazon Neptune

    • A. Amazon OpenSearch Service provides vector search capabilities and is a supported vector store for Bedrock knowledge bases, making it correct for this scenario.
    • B. Amazon Aurora, through its PostgreSQL-Compatible edition with the pgvector extension, is a supported vector store for storing and querying embeddings in a Bedrock knowledge base, making it correct.
    • C. Amazon Neptune can store and query embeddings as a vector store, including graph-based retrieval through Neptune Analytics, making it correct for this scenario.
    • D. AWS Glue is a serverless data integration service for ETL jobs and a data catalog. It can prepare the source documents, but it has no vector index or similarity search, and Bedrock Knowledge Bases cannot use it as a vector store.
    • E. Amazon DynamoDB is a key-value and document database. It does not offer vector similarity search and is not a supported vector store for Bedrock Knowledge Bases. It can hold metadata or application data, but it cannot serve as the embeddings store.
    • F. Amazon Elastic File System provides shared NFS file storage. It does not index vectors or run similarity queries, and Bedrock Knowledge Bases does not support it as a vector store. It can hold files, but it cannot serve as the embeddings index.

    Subdomain 3.2: Choose effective prompt engineering techniques.

    24.Which statement correctly distinguishes prompt poisoning from jailbreaking as risks in prompt engineering?

    1. A.Poisoning corrupts the examples or data that shape a model's behavior, while jailbreaking uses a crafted prompt to bypass its safety instructions at runtime.
    2. B.Poisoning uses a crafted prompt to bypass a model's safety instructions at runtime, while jailbreaking corrupts the examples or data that shape its behavior.
    3. C.Poisoning targets a model's weights and training infrastructure directly, while jailbreaking makes a model disclose its confidential system prompt to a user.
    4. D.Poisoning makes a model disclose its confidential system prompt to a user, while jailbreaking targets its weights and training infrastructure directly.
    Show answer & explanation

    Correct answer: A — Poisoning corrupts the examples or data that shape a model's behavior, while jailbreaking uses a crafted prompt to bypass its safety instructions at runtime.

    • A. Prompt poisoning tampers with the examples, few-shot samples or data that shape how a model behaves. Jailbreaking is a crafted input sent at runtime that gets the model to ignore its built-in safety instructions.
    • B. This swaps the two definitions. Poisoning is the corruption of the examples or data that shape a model, not a runtime attempt to bypass safety instructions. Jailbreaking is the runtime bypass attempt, not a corruption of data.
    • C. Poisoning is not defined by access to weights or training infrastructure, since it works by corrupting the examples or data that shape the model. Jailbreaking is a safety bypass, whereas making a model disclose its system prompt is prompt leaking.
    • D. Disclosing a confidential system prompt is prompt leaking, a separate risk, and poisoning means corrupting examples or data. Jailbreaking needs no access to weights or training infrastructure, because it works through crafted prompts at runtime.

    Subdomain 3.3: Describe the training and fine-tuning process for FMs.

    25.A financial services firm takes a general-purpose foundation model and further trains it on a large collection of unlabeled financial filings and earnings call transcripts so it becomes fluent in financial terminology and context, without targeting one specific downstream task. Which fine-tuning method does this describe?

    1. A.Transfer learning: starts from a pretrained model and trains it on a labeled dataset for a new target task, reusing the representations it already learned.
    2. B.Domain adaptation: continues training on a large body of text from one field, so the model absorbs that field's vocabulary, phrasing and typical context.
    3. C.Instruction tuning: trains on curated prompt and response pairs covering many task types, so the model follows written directions more reliably.
    4. D.Reinforcement learning from human feedback: trains a reward model on human rankings of outputs, then optimizes the model toward preferred responses.
    Show answer & explanation

    Correct answer: B — Domain adaptation: continues training on a large body of text from one field, so the model absorbs that field's vocabulary, phrasing and typical context.

    • A. Transfer learning of this kind aims at one new target task and relies on a labeled dataset for it. The firm has only unlabeled filings and transcripts and no specific downstream task, so this does not describe what it is doing.
    • B. Domain adaptation continues training a general-purpose model on a large body of text from one field. The model gains command of that field's terminology and context without being aimed at a particular task, which matches the firm's approach.
    • C. Instruction tuning needs curated prompt and response pairs, and its goal is better adherence to directions. The firm has unlabeled documents with no such pairs, and its goal is financial fluency, not instruction following.
    • D. This method needs human rankings of model outputs to train a reward model, and it shapes behavior toward preferred responses. The firm has only raw filings and transcripts, with no preference data and no output-quality objective.

    Subdomain 3.4: Describe methods to evaluate FM performance.

    26.A team built a multi-step agent workflow that calls tools and APIs to complete customer requests. Which approaches are appropriate for evaluating this application's performance? (Select TWO.)(Select 2)

    1. A.Task success rate, tracking the share of customer cases the agent closes with the right outcome.
    2. B.Human trace review, examining the agent's reasoning steps and tool calls for errors and risks.
    3. C.ROUGE-L scoring, comparing the agent's final message with a reference answer for text quality.
    4. D.CloudWatch billing metrics, tracking the AWS account's monthly spend across the services it uses.
    5. E.Vector index size, tracking the storage footprint of the knowledge base the agent retrieves from.
    Show answer & explanation

    Correct answers: A, B — Task success rate, tracking the share of customer cases the agent closes with the right outcome.; Human trace review, examining the agent's reasoning steps and tool calls for errors and risks.

    • A. Correct. The share of customer cases that end in the right outcome shows whether the whole workflow works: planning, tool selection and API calls combined. For an agent that takes actions, that end result is the most direct performance measure.
    • B. Correct. A multi-step agent makes chains of decisions and tool calls that automated text metrics cannot judge for correctness or safety. Human raters reading the reasoning steps and tool calls can see whether the right tools were chosen, parameters were valid and unsafe actions were avoided.
    • C. Incorrect. ROUGE-L measures word-sequence overlap with a reference text, so it rates only the wording of the final message. It cannot tell whether the agent called the right tools or APIs or resolved the request. A correct resolution phrased differently would score poorly, and a fluent message from a failed run could score well.
    • D. Incorrect. Billing metrics report account-level monthly spend across all services. They are not tied to individual customer requests, so they say nothing about whether the agent resolved them or used its tools correctly. Cost can be a secondary operational signal, but these totals do not evaluate task performance.
    • E. Incorrect. Index size describes how much storage the knowledge base uses, which is an infrastructure property. It reveals nothing about whether the agent resolved customer requests, reasoned soundly or called its tools correctly. It also does not measure retrieval quality.

    Subdomain 3.2: Choose effective prompt engineering techniques.

    27.A company wants to reduce the risk that users of its Amazon Bedrock chatbot extract confidential internal details or override the assistant's instructions with crafted input. Which TWO actions should the company take? (Select TWO.)(Select 2)

    1. A.Lower the temperature inference parameter to a minimal value, making the model's output consistent and its instruction-following steadier.
    2. B.Remove API keys and customer records from the system prompt, fetching that data through backend calls limited by IAM permissions.
    3. C.Switch to a larger foundation model in Bedrock, giving the chatbot stronger reasoning to recognise and refuse crafted override attempts.
    4. D.Enable the prompt attack content filter in Amazon Bedrock Guardrails, screening user input for jailbreak and injection attempts.
    5. E.Add more few-shot examples of normal conversations to the system prompt, showing the model how to handle unusual user requests.
    Show answer & explanation

    Correct answers: B, D — Remove API keys and customer records from the system prompt, fetching that data through backend calls limited by IAM permissions.; Enable the prompt attack content filter in Amazon Bedrock Guardrails, screening user input for jailbreak and injection attempts.

    • A. Temperature only controls how random token selection is. A low value makes the model more predictable but does not stop it from revealing prompt contents or following instructions a user injects. It is a generation setting, not a security control.
    • B. Correct because anything placed in a prompt can be exposed by crafted queries. Keeping secrets and customer data out of the system prompt, and fetching them through backend calls restricted by IAM permissions, limits what a successful extraction can reveal.
    • C. Model size does not remove prompt injection or jailbreak risk. Larger models can still follow crafted instructions, and a bigger model does nothing to keep confidential details out of the prompt. It is not a security control.
    • D. Correct because the Bedrock Guardrails prompt attack filter is built to detect jailbreak and prompt injection attempts in user input, so crafted override attempts can be screened before the model acts on them.
    • E. Few-shot examples shape format and tone, but they do not prevent user text from overriding instructions. They also make the prompt longer, which adds more content that crafted queries could extract.

    Domain 4: Guidelines for Responsible AI

    Subdomain 4.1: Explain the development of AI systems that are responsible.

    28.A classification model achieves 99% accuracy on its training data but only 61% accuracy on new, unseen validation data. Which effect of bias and variance does this describe?

    1. A.Overfitting
    2. B.Underfitting
    3. C.Fairness violation
    4. D.Data poisoning
    Show answer & explanation

    Correct answer: A — Overfitting

    • A. A large gap between near-perfect training accuracy and much lower validation accuracy is the classic signature of overfitting, where the model memorizes training data rather than generalizing.
    • B. Underfitting shows up as poor performance on both training and validation data, not the high training accuracy seen here.
    • C. A fairness violation involves unequal treatment across demographic groups, which is a different concept from a training-versus-validation accuracy gap.
    • D. Data poisoning refers to deliberately corrupted training data causing malicious behavior, not a natural generalization gap between training and validation accuracy.

    Subdomain 4.1: Explain the development of AI systems that are responsible.

    29.A retail company's customer service chatbot, built on Amazon Bedrock, sometimes returns off-topic or toxic replies and occasionally leaks fragments of customers' personal contact details back into the conversation. The company wants a managed, configurable capability that filters harmful content and redacts sensitive information without retraining the underlying foundation model. Which AWS capability should they configure?

    1. A.Amazon Bedrock Guardrails
    2. B.Amazon Bedrock Agents
    3. C.SageMaker Model Monitor
    4. D.Amazon Augmented AI (A2I)
    Show answer & explanation

    Correct answer: A — Amazon Bedrock Guardrails

    • A. Amazon Bedrock Guardrails provides configurable content filters, denied topics, word filters, and sensitive information filters that can block or mask PII and harmful content without changing the underlying model.
    • B. Amazon Bedrock Agents lets a foundation model plan multi-step tasks and call APIs and knowledge bases on a user's behalf. It does not provide content filters or sensitive-information redaction for the chatbot's replies. Toxic text and leaked contact details would still pass through unless a separate safeguard is applied.
    • C. SageMaker Model Monitor tracks data quality, model quality, and bias drift for deployed SageMaker AI endpoints, not conversational content safety in a chatbot.
    • D. Amazon Augmented AI routes low-confidence predictions to human reviewers and does not filter chat content or redact sensitive information in real time.

    Subdomain 4.2: Recognize the importance of transparent and explainable models.

    30.An insurance company must choose between two models for pricing decisions: a linear model that adjusters can fully explain to customers, and a gradient-boosted ensemble that scores 5% higher on accuracy but cannot be easily explained. Which factor should most directly guide this decision when regulations require customers to receive a clear reason for their premium?

    1. A.How readily each model's premium for one policyholder can be traced to specific rating inputs they can see
    2. B.How much pricing error each model leaves on held-out policies when scored against actual claims experience
    3. C.How quickly each model returns a premium at quote time when customers request prices through the online portal
    4. D.How much effort each model needs to monitor and retrain as claims patterns drift across the policy book
    Show answer & explanation

    Correct answer: A — How readily each model's premium for one policyholder can be traced to specific rating inputs they can see

    • A. Regulations require customers to receive a clear reason for their premium, so the deciding factor is whether an individual premium can be traced to inputs the customer can be shown. The linear model meets this obligation, and a small accuracy gain from a model that cannot do so does not outweigh it.
    • B. Held-out error is a legitimate modelling metric and the ensemble scores better on it, but it says nothing about whether a customer can be given a clear reason for a premium. A mandatory explanation duty outranks a 5% accuracy gain.
    • C. Quote-time latency matters for the customer experience of the portal, but it is an operational property. It does not show whether either model can supply the clear reason for each premium that regulations require.
    • D. Monitoring and retraining effort is a real ongoing cost of running a pricing model, but it is a maintenance concern. It does not determine whether premiums can be explained to customers, which is the binding requirement.

    Subdomain 4.2: Recognize the importance of transparent and explainable models.

    31.Which of the following reflect human-centered design principles for explainable AI? (Select TWO.)(Select 2)

    1. A.Adapt the level of detail and the presentation of an explanation to the background of the person reading it
    2. B.Gather ratings from end users on whether an explanation helped them understand the decision it accompanied
    3. C.Standardize on a single explanation format across the product, for example a feature-importance bar chart
    4. D.Keep user-facing explanations brief and confident by leaving out the model's known limitations and error rates
    5. E.Expose the model's raw internal values, such as weights and activations, in the explanation view for end users
    Show answer & explanation

    Correct answers: A, B — Adapt the level of detail and the presentation of an explanation to the background of the person reading it; Gather ratings from end users on whether an explanation helped them understand the decision it accompanied

    • A. Human-centered explainable AI fits the depth and format of an explanation to its audience. A regulator, a clinician and a customer need different levels of detail to understand the same decision.
    • B. Asking end users whether an explanation helped them understand a decision is a core human-centered practice. Their answers show whether people actually comprehend the explanations, and the interface can be improved from that evidence.
    • C. A single fixed format, even a familiar chart, assumes all audiences read explanations the same way. Data scientists, domain experts and lay users differ in expertise and needs, so one format does not reflect human-centered design.
    • D. Leaving out known limitations and error rates makes explanations look more certain than the model warrants. People then cannot judge how far to rely on a decision, which weakens the trust and transparency that human-centered design aims for.
    • E. Raw weights and activation values are technically faithful but unreadable to most end users. Human-centered design judges an explanation by whether the recipient can understand and act on it, and this does not meet that test.

    Domain 5: Security, Compliance, and Governance for AI Solutions

    Subdomain 5.1: Explain methods to secure AI systems.

    32.A company wants to protect the confidentiality of the datasets it stores for model training and of the traffic it sends to a managed AI service. Which combination correctly describes standard AWS mechanisms for encryption at rest and encryption in transit?

    1. A.Use AWS KMS keys to encrypt stored data and TLS to encrypt traffic between clients and the service
    2. B.Use TLS certificates to encrypt stored data and AWS KMS keys to encrypt traffic between clients and the service
    3. C.Use IAM policies to encrypt stored data and security group rules to encrypt traffic between clients and the service
    4. D.Use AWS KMS keys to encrypt stored data and VPC network ACLs to encrypt traffic between clients and the service
    Show answer & explanation

    Correct answer: A — Use AWS KMS keys to encrypt stored data and TLS to encrypt traffic between clients and the service

    • A. AWS KMS keys are the standard way to encrypt data at rest, for example in S3 or EBS, and TLS is the standard protocol that encrypts data in transit to AWS service endpoints. Each mechanism matches the layer it protects, so this combination covers both the stored training datasets and the traffic sent to the managed AI service.
    • B. This swaps the two mechanisms. TLS certificates secure a network connection between two endpoints and do not encrypt objects or volumes sitting in storage. KMS keys encrypt stored data and do not secure the connection while data moves to the service, so neither requirement is met.
    • C. IAM policies control who may access stored data. They do not encrypt it, so the datasets stay unencrypted at rest. Security group rules filter traffic by port, protocol and source, and they do not encrypt it, so data sent to the service is not protected by this mechanism. Both are access controls, and neither provides encryption.
    • D. KMS keys do correctly cover encryption at rest. VPC network ACLs only allow or deny traffic by address, port and protocol at the subnet level, and they do not encrypt anything. Traffic to the managed AI service therefore still needs TLS to be protected in transit, so the in-transit requirement is not met.

    Subdomain 5.1: Explain methods to secure AI systems.

    33.A data engineering team is setting up the pipeline that will feed customer data into a model training job. Which of the following practices align with secure data engineering best practices for AI systems? (Select TWO.)(Select 2)

    1. A.Anonymize or tokenize sensitive fields in the customer dataset before a training job reads the data from storage
    2. B.Give data scientists administrator access to the production data lake, then remove it once the model is trained
    3. C.Scope IAM policies to specific datasets and roles, and give each training job access only to the data it reads
    4. D.Keep raw training data in a single S3 bucket with a bucket policy that grants read access to the whole organization
    5. E.Run quality checks on internal customer data only, and accept datasets from well-known research groups as they are
    Show answer & explanation

    Correct answers: A, C — Anonymize or tokenize sensitive fields in the customer dataset before a training job reads the data from storage; Scope IAM policies to specific datasets and roles, and give each training job access only to the data it reads

    • A. Correct. Customer data headed for model training should have its sensitive fields anonymized or tokenized first, so the training job and the model never handle raw personal data. This is a standard privacy-enhancing practice for AI data pipelines.
    • B. Incorrect. Administrator rights on the production data lake break least privilege, even if they are removed later. While the rights are held, any data scientist can read, change or delete the customer data and the other datasets in the lake. Secure pipelines grant access scoped to the datasets and roles each person needs.
    • C. Correct. IAM policies scoped to specific datasets and roles apply least privilege to the training data. Each person and job can reach only the data it needs, which limits exposure of the customer data.
    • D. Incorrect. A bucket policy that grants read access to the whole organization lets every principal in it read raw customer data, whether or not it has a reason to. One broad policy replaces per-job and per-role scoping, so it does not meet least-privilege access control for sensitive training data.
    • E. Incorrect. A dataset's origin does not guarantee its quality or integrity. Published datasets can contain errors, bias or poisoned records, so accepting them unchecked leaves a weakness in the training pipeline. Secure data engineering assesses data quality and provenance for every source.

    Subdomain 5.2: Recognize governance and compliance regulations for AI systems.

    34.A marketing team subscribes to a third-party SaaS application that has generative AI features built in, and the team has no ability to customize the underlying model or access its training data. According to the Generative AI Security Scoping Matrix, which perspective should the organization adopt when assessing risk for this use case?

    1. A.Buyer's perspective, focused on reviewing the vendor's data handling terms and enterprise agreement
    2. B.Builder's perspective, focused on threat modeling the model architecture and inference endpoints
    3. C.Builder's perspective, focused on validating the training data used to build the application's model
    4. D.Buyer's perspective, focused on patching and hardening the infrastructure that hosts the model
    Show answer & explanation

    Correct answer: A — Buyer's perspective, focused on reviewing the vendor's data handling terms and enterprise agreement

    • A. A subscribed SaaS application with built-in generative AI, where the organization cannot customize the model, falls in the buy scopes of the matrix. There the organization takes a buyer's perspective and manages risk mainly through the vendor's data handling terms and the enterprise agreement.
    • B. Threat modeling a model architecture and its inference endpoints belongs to a builder's perspective, which applies when the organization builds or hosts its own model. The marketing team only subscribes to a vendor's application and cannot customize the model, so it has no architecture or endpoints to model.
    • C. Validating training data is a builder's task for organizations that fine-tune or train models. The marketing team has no access to the vendor's training data, so it cannot validate that data.
    • D. The buyer's perspective is the right one, but its focus is wrong. The vendor owns and operates the model's hosting infrastructure in a SaaS offering, so the customer cannot patch or harden it. The customer's leverage is contractual and data governance review.

    Subdomain 5.2: Recognize governance and compliance regulations for AI systems.

    35.Which of the following are commonly described as core elements of a data governance strategy for AI systems? (Select TWO.)(Select 2)

    1. A.Defining retention and lifecycle rules for the training and inference data a model uses.
    2. B.Sizing GPU instances and cluster capacity for the jobs that process the training data.
    3. C.Setting residency controls for the regions that store and process the model's data.
    4. D.Tuning temperature and top-p values that shape how outputs vary for the same input.
    5. E.Choosing the orchestration platform that hosts and autoscales the model's endpoints.
    Show answer & explanation

    Correct answers: A, C — Defining retention and lifecycle rules for the training and inference data a model uses.; Setting residency controls for the regions that store and process the model's data.

    • A. Correct. Retention and lifecycle rules decide how long training and inference data is kept and when it is archived or deleted. They are a core data governance element because they support privacy, compliance and accountability for the data an AI system uses.
    • B. Incorrect. Sizing GPU instances and cluster capacity is a compute capacity and cost decision, even when the jobs it serves process training data. It does not set how that data is managed, protected, retained or located, so it is not a data governance element.
    • C. Correct. Residency controls fix the regions where data may be stored and processed, which matters for regulatory and contractual obligations. This makes them a core data governance element for AI systems.
    • D. Incorrect. Temperature and top-p are inference parameters that control the randomness and variety of generated outputs. They change model behavior at response time and do not govern how data is managed, retained or located.
    • E. Incorrect. The orchestration platform that hosts and scales model endpoints is an infrastructure and operations decision. It does not control how data is managed, retained or located, so it is not a data governance element.

    Want the full experience?

    These are just samples. Practice the full AWS Certified AI Practitioner (AIF-C01) question bank in quiz mode — free, no signup, with domain practice and exam simulation.