AWS Certified AI Practitioner (AIF-C01) Practice Exam

What the AWS Certified AI Practitioner (AIF-C01) exam covers

• Fundamentals of AI and ML176 questions
• Fundamentals of generative AI211 questions
• Applications of foundation models246 questions
• Guidelines for responsible AI123 questions
• Security, compliance, and governance for AI solutions124 questions

Free AWS Certified AI Practitioner (AIF-C01) sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 880.

1. Question 1Fundamentals of AI and ML

   A manufacturer has trained a defect detection model in SageMaker and now needs to run that model on a factory floor camera device with limited compute and no reliable connection to the cloud, so the model must be compiled and optimized specifically for that device's hardware and runtime. Which SageMaker capability is designed for this optimization?

   • ASageMaker Autopilot, for automatically tuning candidate models
   • BSageMaker Ground Truth, for generating labeled defect images
   • CSageMaker Neo, which compiles models for specific edge hardwareCorrect
   • DSageMaker Pipelines, for orchestrating multi-step training workflows
   ✓ Correct answer: C

   Neo takes a trained model and compiles it into an optimized executable tuned for a chosen target, such as a particular processor or edge device runtime, improving inference speed and reducing the memory footprint, which suits deploying a model to a camera device with limited compute and no guaranteed cloud connectivity.

   Why the other options are wrong

   • AAutopilot automates algorithm selection and hyperparameter tuning during model development, it does not compile a finished model for a specific device.
   • BGround Truth is a data labeling service for creating training datasets, it has no role in compiling a trained model for edge hardware.
   • DPipelines orchestrates the sequence of steps in a training and deployment workflow, it does not perform hardware-specific model compilation.
2. Question 2Fundamentals of AI and MLSelect all that apply

   Which TWO statements correctly describe the relationship between artificial intelligence (AI), machine learning (ML), deep learning (DL), and generative AI?

   • ADeep learning is a subset of machine learning that uses multi-layered neural networksCorrect
   • BArtificial intelligence is a subset of deep learning focused narrowly on neural networks
   • CMachine learning is a subset of artificial intelligence in which systems learn from dataCorrect
   • DGenerative AI can only be implemented using unsupervised learning techniques
   ✓ Correct answer: A, C

   Deep learning is a specialized branch of machine learning characterized by deep, multi-layered neural network architectures. Machine learning itself sits within the broader discipline of artificial intelligence, encompassing techniques where systems learn from data rather than following purely hand-coded rules. These nesting relationships, AI containing ML which contains DL, are foundational to understanding AI terminology.

   Why the other options are wrong

   • BThis reverses the actual hierarchy; artificial intelligence is the broadest term, and deep learning is one of its narrower subsets, not the other way around.
   • DGenerative AI models are typically built using deep learning architectures and can be trained using various techniques, including self-supervised and supervised learning, not exclusively unsupervised learning.
3. Question 3Fundamentals of generative AI

   Two different foundation models may split the exact same sentence into different numbers of tokens. What best explains this variation?

   • AOnly the sentence's character length determines token count, and every model computes it identically
   • BTokens are randomly reassigned to different words every single time the text is processed
   • CEach model may use a different tokenizer and vocabulary, so it segments text differentlyCorrect
   • DEach model runs on a different amount of allocated RAM and compute resources during inference
   ✓ Correct answer: C

   Different models are trained with different tokenizers and vocabularies, for example different byte pair encoding merges, so the same text can be split into different numbers and boundaries of tokens across models. This means token counts for the same text are model-specific, not universal.

   Why the other options are wrong

   • ACharacter length alone does not determine token count identically across models, since tokenizers differ in how they segment words and subwords.
   • BTokenization is deterministic given a fixed tokenizer and vocabulary, not random from one run to the next.
   • DThe amount of RAM or compute allocated during inference is an infrastructure detail unrelated to how a specific tokenizer segments text.
4. Question 4Fundamentals of generative AI

   Which statement BEST describes a common tradeoff when choosing a larger, more capable foundation model instead of a smaller one for a given task?

   • ALarger models always require significantly less labeled data to fine-tune successfully
   • BLarger models remove the need for any prompt engineering work during deployment
   • CLarger models almost always have a lower cost per token than smaller models
   • DLarger models typically produce higher-quality outputs but at higher inference cost and latencyCorrect
   ✓ Correct answer: D

   A well-known tradeoff in foundation model selection is that larger, more capable models tend to produce higher-quality results on complex tasks, but they generally require more compute per request, which raises both inference cost and latency compared to smaller models.

   Why the other options are wrong

   • AFine-tuning data needs depend on the task and approach, not simply on model size; larger models do not inherently need less data.
   • BPrompt engineering remains valuable regardless of model size to get the best results.
   • CLarger models generally cost more per token to run due to greater compute requirements, not less.
5. Question 5Fundamentals of generative AI

   A company evaluating Amazon Bedrock for a customer support use case asks whether their prompts and outputs will be used to train the underlying foundation models. What is the accurate general statement about Amazon Bedrock's approach to customer data?

   • AAmazon Bedrock does not use customer inputs and outputs to train the underlying base models, and customer content is not shared with model providers for that purposeCorrect
   • BAmazon Bedrock has no documented data handling policy regarding customer prompts, leaving customers to assume how their submitted content might be used or stored
   • CAmazon Bedrock requires customers to opt into model training as a mandatory condition of use, and declining that option blocks access to foundation models
   • DAmazon Bedrock always uses customer prompts and outputs to retrain the base foundation models by default, unless a customer explicitly requests an exception
   ✓ Correct answer: A

   Amazon Bedrock is designed so that customer prompts and generated outputs are not used to train the underlying base foundation models, and this content is not shared with third party model providers for improving their general purpose models, which helps address data privacy concerns for regulated workloads. This is a documented AWS commitment intended to give enterprises confidence when sending business data into Bedrock hosted models. Encryption and access controls further support secure use of customer data within the service.

   Why the other options are wrong

   • BAmazon Bedrock does have documented data handling and privacy commitments, so customers are not left to guess; claiming there is no policy is inaccurate.
   • CThere is no mandatory opt in to model training as a condition of using Amazon Bedrock, and no such access block exists; contributing data to training is not required.
   • DThis is the opposite of Amazon Bedrock's actual approach; base model retraining on customer data by default, with or without an exception process, is not how the service operates.
6. Question 6Applications of foundation modelsSelect all that apply

   Which of the following are vector-capable data stores that can serve as the underlying vector store for a Retrieval Augmented Generation solution on AWS? (Select TWO.)

   • AAmazon DynamoDB with on-demand capacity mode
   • BAmazon OpenSearch ServiceCorrect
   • CAmazon Aurora PostgreSQL with pgvectorCorrect
   • DAmazon Redshift Serverless data warehouse
   ✓ Correct answer: B, C

   Amazon OpenSearch Service offers a k-NN plugin for vector indexing and search, and Amazon Aurora PostgreSQL supports the pgvector extension for storing and querying embedding vectors with SQL. Both are suitable as the vector store behind a RAG retrieval pipeline.

   Why the other options are wrong

   • AAmazon DynamoDB, even with on-demand capacity mode, is a fully managed key-value and NoSQL store; it does not provide native vector similarity indexing and is not one of the vector store options for Bedrock Knowledge Bases.
   • DAmazon Redshift Serverless is a managed data warehouse optimized for large-scale analytical SQL queries; it is not offered as a supported vector store for Bedrock Knowledge Bases.
7. Question 7Applications of foundation models

   Approximate nearest neighbor (ANN) algorithms such as HNSW are widely used in vector search because they trade off which two factors?

   • AIAM policy rule complexity versus the total number of allowed users
   • BModel temperature setting versus maximum output token length allowed
   • CEncryption key strength versus available network bandwidth throughput
   • DSearch speed versus perfect recall of the exact nearest vectorsCorrect
   ✓ Correct answer: D

   Exhaustive, exact, nearest neighbor search compares a query vector against every stored vector, which becomes slow at scale. ANN algorithms like HNSW build index structures that return very fast results that are usually, but not always guaranteed to be, the true closest matches, trading a small amount of recall accuracy for large gains in query speed.

   Why the other options are wrong

   • AIAM policy complexity and user counts are access control concerns, unrelated to the ANN speed versus accuracy tradeoff.
   • BModel temperature and output token length are generation-time parameters of the foundation model, unrelated to vector index search algorithms.
   • CEncryption key strength and network bandwidth throughput are unrelated to the ANN algorithm design tradeoff.
8. Question 8Applications of foundation models

   A company evaluating customization options for a customer service assistant lists these needs: frequent factual updates from an internal knowledge base, minimal upfront data preparation, and no changes to model weights. Which approach best satisfies all three needs simultaneously?

   • ARetrieval Augmented Generation connected to the knowledge baseCorrect
   • BFull fine-tuning refreshed and rerun against the knowledge base every week
   • CPrompt engineering with static built-in facts hardcoded into the system prompt
   • DContinued pre-training refreshed and rerun against the knowledge base every week
   ✓ Correct answer: A

   RAG satisfies all three stated needs because it retrieves updated content from the knowledge base at query time without any weight changes and without requiring labeled training data preparation, directly matching frequent updates, minimal data prep, and unchanged weights.

   Why the other options are wrong

   • BWeekly full fine-tuning also changes model weights and requires meaningful retraining effort, conflicting with the stated needs.
   • CHardcoding static facts into a prompt cannot reflect frequent knowledge base updates without manual editing every time, failing the freshness requirement.
   • DContinued pre-training requires substantial data preparation and retraining effort, and it does change model weights, conflicting with two of the stated needs.
9. Question 9Guidelines for responsible AI

   A data science team notices that their resume screening model was trained on historical hiring data that underrepresents certain applicant groups. They apply reweighting to give underrepresented examples more influence during training and rerun a disparate impact analysis before deployment. This work most directly supports which responsible AI dimension?

   • AInterpretability
   • BTransparency
   • CFairnessCorrect
   • DPrivacy and security
   ✓ Correct answer: C

   Reweighting underrepresented training examples and running a disparate impact analysis are standard bias mitigation techniques used to improve fairness, since they directly target unequal treatment or outcomes that arise from skewed training data. Fairness practices aim to detect and correct such imbalances before they propagate into production decisions.

   Why the other options are wrong

   • AInterpretability concerns how understandable a model's internal logic is to humans, which is a separate concern from correcting data-driven bias.
   • BTransparency is about disclosing information regarding how a system works, not about the technical process of correcting data imbalance.
   • DPrivacy and security concerns protecting data and models from misuse or unauthorized access, not adjusting training data to reduce disparate outcomes.
10. Question 10Security, compliance, and governance for AI solutions

    A company deployed a SageMaker real-time inference endpoint and wants to ensure that only authenticated, authorized AWS principals within its VPC can invoke it, with no public endpoint exposure. Which approach addresses this?

    • ADeploy the endpoint with VPC configuration for network isolation and require IAM authentication (SigV4) for InvokeEndpoint calls, using a private interface VPC endpoint for the SageMaker runtimeCorrect
    • BPublish the endpoint through a public Amazon API Gateway deployment secured only with a static API key, without placing any resources inside the company's private VPC
    • CRely on the default SageMaker endpoint configuration without making any additional changes, since SageMaker inference endpoints are assumed to be private and authenticated by default
    • DShare the endpoint's invocation URL with the internal team only through internal email and a wiki page, without applying any additional authentication or network controls
    ✓ Correct answer: A

    Placing the endpoint's model containers in the customer's VPC, requiring IAM SigV4-signed requests for InvokeEndpoint, and using a private interface VPC endpoint for the sagemaker-runtime API together ensure invocation traffic stays private and only authorized IAM principals can call the endpoint.

    Why the other options are wrong

    • BExposing the endpoint through a public API Gateway with only a static API key reintroduces public internet exposure and weak authentication, which directly contradicts the requirement.
    • CSageMaker runtime API calls are reachable over the public sagemaker-runtime endpoint by default unless the customer explicitly configures VPC isolation, private connectivity, and IAM authentication; assuming default privacy is incorrect.
    • DSharing a URL through email or a wiki is not an access control mechanism at all, provides no authentication, and is trivially insecure regardless of how internally it is distributed.

AWS Certified AI Practitioner (AIF-C01) practice exam FAQ