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

Correct answer: A — Computer vision, because it analyzes pixel data in images to detect visual defects and anomalies automatically.

• A. This is correct because computer vision is the AI field that analyzes and interprets visual data such as photographs, making it suited to detecting scratches, missing parts, and misalignment in images.
• B. This is incorrect because the scenario is based on analyzing photographs, not written technician notes, so a text-focused capability does not address the task.
• C. This is incorrect because the manufacturer needs to flag defects in each individual photo as it is captured, not forecast an aggregate defect rate over a future time period.
• D. This is incorrect because the input described is image data from cameras, not audio recordings, so converting speech to text does not apply here.

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.

Correct answer: A — The task requires a deterministic outcome, and ML models produce probabilistic predictions instead of guaranteed exact results

• A. Balance arithmetic is a deterministic rule-based calculation with one correct answer, and ML models instead output probabilistic predictions that approximate an outcome rather than guarantee it.
• B. ML models routinely process numeric and financial data, such as fraud scores or credit risk features, so this is not the limiting factor here.
• C. Storage capacity is unrelated to why ML is unsuitable; the mismatch is about deterministic versus probabilistic outputs, not resource limits.
• D. ML models regularly work with structured, tabular data such as transaction records, so this claim about text-only processing is inaccurate.

Correct answer: A — Experimentation, the practice of iterating on models and configurations to find the best candidate

• A. Trying multiple algorithms and hyperparameters while tracking metrics for comparison is the definition of experimentation, an MLOps concept practiced before a model is finalized.
• B. Model monitoring applies to a model already running in production, not to comparing candidate models before one has been selected.
• C. Technical debt management addresses accumulated shortcuts in code or process, which is not what is being described when comparing candidate models.
• D. Production readiness is a certification step that happens after a model is chosen, not the process of comparing multiple candidates.

Correct answer: A — Real-time inferencing, because it returns predictions for individual requests with very low latency as they arrive.

• A. Real-time inferencing is correct because it is designed to serve predictions for single incoming requests with millisecond-scale latency, matching the airline's need for an instant offer.
• B. Batch inferencing processes large groups of records together on a schedule and is not suited to returning a prediction the moment a single customer loads a page.
• C. There is no inferencing mode called 'training inferencing'; training adjusts model weights and is a separate phase that happens before inferencing, not on every page load.
• D. Offline or overnight processing introduces delay that does not meet a millisecond response requirement for an individual browsing customer.

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 Polly converts written text into lifelike spoken audio and has no capability for analyzing image or video content.

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. Development cost is incorrect because it is a business metric reflecting spending on building the model, not a measure of classification performance.
• D. Customer feedback score is incorrect because it reflects subjective user satisfaction rather than the statistical balance of precision and recall for a classifier.

Correct answer: D — Prompt engineering, which crafts and refines instructions and examples to guide the model's output

• A. Fine-tuning changes the model's underlying weights through additional training, but the scenario explicitly avoids retraining and only adjusts the instructions given to the model.
• B. Distillation trains a separate, smaller model to mimic a larger one's behavior; it is a training technique, not a way of rewording instructions to an existing deployed model.
• C. Data selection concerns choosing datasets before pre-training begins, which has no connection to adjusting a chatbot's instructions at inference time.
• D. This is correct: prompt engineering refines the wording, examples, and structure of instructions given to a model to steer its output without changing its weights.

Correct answer: B — Cost is calculated based on the number of input and output tokens processed for each request

• A. A flat, usage-independent monthly fee describes a subscription model, not the per-token metering that most foundation model inference pricing actually uses.
• B. This is correct: token-based pricing charges based on the number of input tokens sent and output tokens generated, so longer prompts and responses cost more.
• C. Charging strictly per API call regardless of size would ignore the actual amount of text processed, which is not how token-based pricing works.
• D. The size of the model's stored weight files reflects infrastructure hosting costs, not the per-request charge tied to the tokens processed during inference.

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. The number of model parameters is a technical architecture detail that does not measure business value or cost-effectiveness.
• C. Vector embedding dimensionality is a technical property of how text is represented internally and has no direct link to cost justification.
• D. Token vocabulary size describes how many distinct tokens a model recognizes, which does not indicate whether the tool delivers business value relative to cost.

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.

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.

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. Fine-tuning refers to further training a foundation model's weights on labeled examples, which is a separate step from breaking documents into passages.
• C. Tokenization is the process of breaking text into the smaller units (tokens) a model uses internally to process input, which is a different and more granular operation than splitting a manual into paragraph-level passages.
• D. Diffusion describes an image-generation technique that iteratively removes noise to create pictures, which has no relationship to document preprocessing for a knowledge base.

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. Multi-modality refers to a model's ability to process multiple input or output types such as text and images, which is unrelated to generating a false citation.

Correct answer: B — Amazon SageMaker JumpStart

• A. Bedrock Data Automation extracts structured insights from unstructured multimodal content such as documents and images, and does not provide a catalog of pretrained models to fine-tune within 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 Q Developer is a coding assistant that helps write and explain code, not a catalog of pretrained models for fine-tuning inside SageMaker Studio.
• D. Bedrock Guardrails applies content filtering policies to model input and output, and is unrelated to browsing or fine-tuning pretrained models in SageMaker Studio.

Correct answer: A — Applications inherit AWS data center security controls, encryption capabilities, and compliance certifications that support responsible and safe AI deployments.

• A. This is correct because building on AWS infrastructure lets applications inherit data center security controls, encryption options, and compliance certifications that support responsible, safe, and secure GenAI deployments.
• B. AWS infrastructure provides tools and certifications that help with compliance, but customers still share responsibility for configuring their applications to meet specific regulatory requirements.
• C. AWS infrastructure does not guarantee bias-free or fully accurate model outputs; hallucinations and bias remain risks that require responsible AI practices to manage.
• D. AWS infrastructure does not provide unlimited free compute; customers pay for the compute and inference resources they consume.

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.

Correct answer: A — Responses become more random and varied

• A. Raising the temperature parameter increases randomness in token selection, producing more varied and creative outputs, which suits generating diverse marketing taglines.
• B. Temperature controls output randomness, not output length, so raising it does not shorten the generated text.
• C. A higher temperature makes outputs less deterministic and less repetitive, which is the opposite of this effect.
• D. Temperature has no effect on which languages a model can generate; that depends on the model's trained language capabilities.

Correct answer: C — Prompt poisoning, where malicious or corrupted examples are inserted into the prompt to bias the model's output

• A. Exposure is about instructions being leaked outward to users, not about examples being altered inward by an insider. The direction and mechanism of this scenario do not match exposure.
• B. Hijacking happens through a user's runtime message overriding instructions during a conversation, not through pre-deployment edits to stored few-shot examples. This scenario occurred before any user interaction.
• C. Poisoning describes exactly this situation: corrupted or biased example pairs were inserted into the prompt template itself, causing the model's future outputs to reflect that bias. The unauthorized edit to stored examples is the defining characteristic of poisoning.
• D. Jailbreaking involves a user's runtime input crafted to bypass safety restrictions, not an unauthorized edit made to a stored prompt template. No safety guardrail was bypassed in this scenario.

Correct answer: B — A lack of representativeness in the training data across languages and regions

• A. Labeling accuracy affects whether examples have correct target outputs, but the described failure — poor performance on non-English and regional queries — stems from what languages and regions were included, not from mislabeled examples.
• B. A training set drawn almost entirely from one language and region lacks representativeness, so the model never learned patterns needed for other languages or regional phrasing, which explains its poor global performance.
• C. Strict governance controls affect data access and retention policies, not which languages or regions are represented in the dataset, so this would not directly cause the described performance gap.
• D. The scenario does not mention any reinforcement learning process being used at all, so RLHF reliance cannot explain a language and region coverage gap in the fine-tuning data.

Correct answer: A — To measure model performance against a standardized, publicly available set of tasks so results can be compared across models.

• A. Standardized public task sets let teams compare a model's performance consistently against other models, which is the core purpose of benchmark datasets.
• B. Benchmark datasets are used for evaluation, not to expand or generate fine-tuning training data.
• C. Encrypting stored outputs is a data protection concern unrelated to the purpose of benchmark-based evaluation.
• D. Benchmark datasets support scoring and comparison; they do not automatically trigger deployment decisions.

Correct answer: A — Whether the retrieved documents are relevant and sufficient to support the generated answer.

• A. Answers that are fluent but unsupported point directly to a retrieval quality problem, so evaluating whether retrieved documents actually back up the generated answer is the right focus.
• B. Response time is a performance concern, not the cause of answers being fluent yet factually unsupported by retrieved content.
• C. Storage cost is a budgeting concern unrelated to why generated answers lack grounding in retrieved documents.
• D. Formatting consistency affects presentation, not whether the retrieved context actually supports the generated answer's claims.

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 Lambda is a serverless compute service used to run code in response to events, such as processing files or invoking APIs, and does not store or index vector embeddings, so it is incorrect.
• E. Amazon CloudFront is a content delivery network that caches and serves content close to users, and it has no capability to store or query vector embeddings, so it is incorrect.
• F. Amazon Simple Notification Service is a pub/sub messaging service for sending notifications between systems, and it does not provide vector storage or similarity search, so it is incorrect.

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

• A. Prompt poisoning corrupts the examples, training data, or few-shot samples that influence a model's behavior, while jailbreaking crafts a runtime prompt designed to bypass the model's built-in safety instructions.
• B. This reverses the two definitions; poisoning targets the data or examples shaping the model, and jailbreaking targets safety instructions at runtime, not the other way around.
• C. Neither risk requires gaining direct access to a model's underlying weights or training infrastructure; both are achieved through manipulated inputs rather than infrastructure access.
• D. Unintentional system prompt exposure describes prompt exposure, a separate risk from poisoning and jailbreaking, which involve corrupting inputs or bypassing safety behavior rather than accidental disclosure.

Correct answer: B — Adapting the model for a specific domain, which further trains it on domain-specific text so it becomes more fluent in that field's language and context.

• A. This reuses a model across an unrelated task and modality, whereas the scenario keeps the same modality and instead builds deeper fluency within one field.
• B. This is correct because further training a general-purpose model on a large body of domain-specific, unlabeled text so it becomes fluent in that field's language is exactly what adapting a model for a specific domain means.
• C. This requires paired instructions and desired responses, but the scenario describes training on unlabeled filings and transcripts with no such paired examples.
• D. This relies on human preference rankings of model outputs as a reward signal, which is not present in a scenario describing further training on unlabeled documents.

Correct answer: A — BLEU, which scores n-gram precision against reference translations with a brevity penalty

• A. BLEU is correct because it is designed for translation evaluation, scoring n-gram precision against reference translations and applying a brevity penalty for short outputs.
• B. ROUGE is incorrect here because it is a recall-oriented metric typically applied to summarization tasks, not translation, and does not include a brevity penalty.
• C. BERTScore is incorrect because it relies on embedding similarity rather than counting matching n-grams or applying a brevity penalty.
• D. F1 score is incorrect because it is a classification metric balancing precision and recall of labels, not a text-generation overlap metric with a brevity penalty.

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. SageMaker Clarify computes bias and explainability metrics for trained models and is not designed to filter live conversational content or redact PII in real time.
• 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.

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.

Correct answer: A — The regulatory requirement for clear, customer-facing reasoning should outweigh the small accuracy gain

• A. When a legal requirement mandates that customers receive clear reasons for a decision, an explainable model that satisfies that obligation should take priority over a modest accuracy improvement from an unexplainable ensemble.
• B. Always choosing the highest-accuracy model ignores the stated regulatory obligation to provide customers with a clear, explainable reason for their premium.
• C. Training speed is an operational convenience that does not address the regulatory need for customer-facing explainability.
• D. Artifact storage size is an infrastructure cost consideration that has no bearing on whether the model can satisfy the explainability regulation.

Correct answer: A — Overfitting, where the model learns patterns specific to the training data that do not generalize

• A. Overfitting is correct because a large gap between very high training accuracy and much lower accuracy on unseen data is the classic sign that the model memorized training-specific patterns instead of learning generalizable ones.
• B. Underfitting is incorrect because it describes a model that performs poorly on both training and unseen data due to being too simple, whereas this model performs extremely well on training data.
• C. Data drift is incorrect because it describes production input data changing after deployment over time, not a one-time gap between training accuracy and accuracy on unseen evaluation data.
• D. Label noise is incorrect because it refers to incorrect labels degrading overall model quality on both training and evaluation data, not a large accuracy gap between the two.

Correct answers: A, B — Tailor the depth and format of an explanation to the knowledge level and needs of the person receiving it; Provide a way for end users to give feedback on whether an explanation actually helped them understand a decision

• A. Human-centered explainable AI design calls for adapting explanation depth and format to the audience, since a regulator, a clinician, and an end user need different levels of detail.
• B. Collecting user feedback on whether an explanation was actually understood is a core human-centered practice, since it lets the explanation interface be improved based on real user comprehension.
• C. Using one fixed explanation format for every user ignores differences in expertise and needs, which runs counter to human-centered design rather than reflecting it.
• D. Hiding known model limitations from users undermines trust and transparency, which is the opposite of what human-centered explainable AI design aims to achieve.
• E. Training time is a model-development concern unrelated to how understandable an explanation is to the person receiving it, so it is not a human-centered design principle.

Correct answer: A — To document where data originated and how it was transformed on its path to the model

• A. Data lineage tracks the origin of data and every transformation it undergoes, giving teams traceability and accountability for what fed into a model.
• B. Data lineage is a traceability and documentation practice; it has no relationship to reducing the storage footprint of data.
• C. Tracking where data came from does not itself change model behavior or improve accuracy; it is a governance and traceability practice.
• D. Lineage tracking documents data movement and transformation history; encryption is a separate data protection control, not a lineage function.

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

• A. For an off-the-shelf application where the organization cannot customize the model, the scoping matrix recommends a buyer's perspective centered on reviewing the vendor's data handling terms and enterprise agreement rather than the underlying technology.
• B. A builder's perspective involving threat modeling applies to more customized or internally developed use cases where the organization controls the model, which does not describe a subscribed third-party SaaS feature.
• C. Securing a training pipeline is only relevant when the organization actually operates that pipeline, which is not the case for a purchased SaaS application with an embedded model.
• D. Patching underlying infrastructure is the vendor's responsibility in an off-the-shelf SaaS scenario, and a buyer in this scope focuses on contractual and data governance review rather than infrastructure patching.

Correct answer: A — Encrypt stored data using AWS KMS-managed keys, and protect data moving over the network using TLS

• Encrypt stored data using AWS KMS-managed keys, and protect data moving over the network using TLS. AWS KMS manages the encryption keys used to protect data at rest in services like S3 and SageMaker, while TLS is the standard protocol used to encrypt data in transit between clients and AWS services.
• Encrypt stored data using TLS certificates, and protect data moving over the network using AWS KMS keys. This swaps the two mechanisms; TLS is the transport-layer protocol for data in transit, and KMS keys are used for encrypting stored data, not the reverse.
• Rely on IAM password policies to encrypt stored data, and use VPC route tables to encrypt network traffic. IAM password policies govern user credential strength and route tables direct network paths; neither mechanism performs encryption of stored data or network traffic.
• Rely on Amazon CloudWatch alarms to encrypt stored data, and use security group rules to encrypt network traffic. CloudWatch alarms monitor metrics and trigger notifications, and security groups act as network traffic filters; neither one applies cryptographic encryption to data at rest or in transit.

Correct answer: A — AWS Config with custom rules and configuration history

• AWS Config with custom rules and configuration history. AWS Config rules continuously evaluate resource settings such as encryption status and security group ports against defined conditions and flag noncompliant resources, while its configuration history lets the team view settings from months earlier. This combination of continuous rule evaluation plus historical lookback matches every part of the requirement.
• AWS Artifact with downloaded compliance reports. AWS Artifact provides static, pre-generated compliance documents about AWS's own infrastructure rather than continuously evaluating a customer's individual S3 buckets or security groups. It offers no mechanism for tracking a customer resource's configuration history.
• Amazon Inspector with vulnerability scan findings. Amazon Inspector focuses on scanning compute resources and container images for software vulnerabilities and network reachability issues, not on evaluating encryption settings against a custom internal policy or storing configuration history. Its findings are vulnerability-oriented rather than configuration-drift-oriented.
• AWS Trusted Advisor with cost optimization checks. Trusted Advisor offers point-in-time best-practice recommendations across several categories but does not continuously evaluate custom internal policies or retain a historical record of a resource's past configuration. It lacks the customizable rule and history features the scenario requires.