Free sampler · unofficial
Free AIF-C01 practice exam questions
Twenty original practice questions for the AWS Certified AI Practitioner (AIF-C01) exam, spread across all five domains in line with the official weights. Each answer explains why every option is right or wrong and quotes the official AWS documentation page that proves it.
Take them as an interactive set in the practice tool, or read them below with the answers hidden until you open them. There is no sign-up and no email.
What the 20 questions cover
| Domain | Official weight | Free questions |
|---|---|---|
| D1: Fundamentals of AI and ML | 20% | 4 |
| D2: Fundamentals of GenAI | 24% | 5 |
| D3: Applications of Foundation Models | 28% | 6 |
| D4: Guidelines for Responsible AI | 14% | 3 |
| D5: Security, Compliance, and Governance for AI Solutions | 14% | 2 |
Every question here passed the same checks as the full bank: the quoted sentence was found on the live official page, and an independent blind reviewer matched the answer. How we check questions.
The questions
Answers are hidden. Open “Show answer and source” under a question after you have picked yours.
Question 1 of 20 · D1: Fundamentals of AI and ML · aif-d1-001
A team wants to describe the AI technique that uses layered neural networks inspired by the human brain to recognize complex patterns in images, text, and sound. Which term describes this technique?
- Option A: Deep learning
- Option B: Rule-based expert systems
- Option C: Linear regression
- Option D: Robotic process automation
Show answer and source
Answer: A. Deep learning models use layers of interconnected nodes (neurons) and can recognize complex patterns in pictures, text, and sounds.
- A (correct): Correct. AWS defines deep learning as an AI method that processes data in a way inspired by the human brain.
- B (incorrect): Expert systems apply hand-written rules; they do not learn layered representations from data.
- C (incorrect): Linear regression is a simple statistical model for numeric prediction, not a layered neural network.
- D (incorrect): RPA scripts repetitive UI tasks; it is not a learning technique.
Official sourceaws.amazon.com
Deep learning is an artificial intelligence (AI) method that teaches computers to process data in a way inspired by the human brain.
aws.amazon.com/what-is/deep-learning/ (official page)Question 2 of 20 · D1: Fundamentals of AI and ML · aif-d1-002
Which field of AI focuses on enabling computers to interpret, manipulate, and comprehend human language?
- Option A: Computer vision
- Option B: Reinforcement learning
- Option C: Time-series forecasting
- Option D: Natural language processing (NLP)
Show answer and source
Answer: D. NLP lets systems classify, sort, filter, and understand intent or sentiment in text and voice data.
- A (incorrect): Computer vision derives information from images and video, not language.
- B (incorrect): Reinforcement learning is a training approach based on rewards, not a language-specific field.
- C (incorrect): Forecasting predicts future numeric values from historical data.
- D (correct): Correct. This is the AWS definition of NLP.
Official sourceaws.amazon.com
Natural language processing (NLP) is technology that allows computers to interpret, manipulate, and comprehend human language.
aws.amazon.com/what-is/nlp/ (official page)Question 3 of 20 · D1: Fundamentals of AI and ML · aif-d1-003
A city wants software that derives meaningful information from traffic-camera video, such as detecting vehicles. Which field of AI does this describe?
- Option A: Natural language processing
- Option B: Speech synthesis
- Option C: Computer vision
- Option D: Recommendation systems
Show answer and source
Answer: C. Computer vision applications process visual data for tasks such as object identification, OCR, facial recognition, and monitoring.
- A (incorrect): NLP works with human language, not visual inputs.
- B (incorrect): Speech synthesis turns text into audio.
- C (correct): Correct. Computer vision derives meaningful information from images, videos, and other visual inputs.
- D (incorrect): Recommenders suggest items based on behavior; they do not analyze video frames.
Official sourceaws.amazon.com
Computer vision is a field of artificial intelligence (AI) that enables computers to derive meaningful information from digital images, videos, and other visual inputs.
aws.amazon.com/what-is/computer-vision/ (official page)Question 4 of 20 · D1: Fundamentals of AI and ML · aif-d1-004
Which statement best describes the relationship between machine learning (ML) and artificial intelligence (AI)?
- Option A: AI is a subset of ML that focuses only on neural networks.
- Option B: ML is a type of AI that performs data analysis tasks without explicit instructions.
- Option C: ML and AI are unrelated fields that share no techniques.
- Option D: ML always requires a human to write a rule for every possible input.
Show answer and source
Answer: B. AI is the broad field; ML is a type of AI that identifies patterns in historical data to make predictions on new data.
- A (incorrect): The relationship is reversed; ML sits inside the broader field of AI.
- B (correct): Correct. AWS describes ML as a type of AI that learns from data rather than explicit instructions.
- C (incorrect): ML is one of the main ways AI is built.
- D (incorrect): Learning patterns from data instead of hand-written rules is the point of ML.
Official sourceaws.amazon.com
Machine learning is a type of artificial intelligence that performs data analysis tasks without explicit instructions.
aws.amazon.com/what-is/machine-learning/ (official page)Question 5 of 20 · D2: Fundamentals of GenAI · aif-d2-001
In the context of foundation models on Amazon Bedrock, what is a token?
- Option A: An API key that authorizes calls to a model
- Option B: A fixed block of exactly 1,000 characters
- Option C: A labeled example used for fine-tuning
- Option D: A sequence of characters that a model can interpret or predict as a single unit of meaning
Show answer and source
Answer: D. Models read and generate text as tokens; for example "-ed" or "?" can each be a token.
- A (incorrect): Credentials are not tokens in the generative AI sense.
- B (incorrect): Tokens vary in length; they are not fixed-size character blocks.
- C (incorrect): A labeled example is a training record, not a token.
- D (correct): Correct. This is the Amazon Bedrock definition; a token can be a word, part of a word, punctuation, or a common phrase.
Official sourcedocs.aws.amazon.com
A sequence of characters that a model can interpret or predict as a single unit of meaning.
docs.aws.amazon.com/bedrock/latest/userguide/key-definitions.html (official page)Question 6 of 20 · D2: Fundamentals of GenAI · aif-d2-002
What are embeddings in machine learning?
- Option A: Numerical representations of real-world objects that capture their properties and relationships
- Option B: Compressed archives of training data
- Option C: Hand-written rules that describe each object
- Option D: Human labels attached to images
Show answer and source
Answer: A. Embeddings convert objects such as words or images into vectors so similar items sit close together.
- A (correct): Correct. AWS defines embeddings as numerical representations of real-world objects.
- B (incorrect): Embeddings are vectors, not file archives.
- C (incorrect): Embeddings are learned automatically, not hand-written.
- D (incorrect): Labels are annotations; embeddings are numeric vectors.
Official sourceaws.amazon.com
Embeddings are numerical representations of real-world objects
aws.amazon.com/what-is/embeddings-in-machine-learning/ (official page)Question 7 of 20 · D2: Fundamentals of GenAI · aif-d2-003
A team stores document embeddings and must quickly find the items most similar to a query embedding. What is the main capability that a vector database adds?
- Option A: Translating documents into other languages
- Option B: Enforcing relational foreign-key constraints only
- Option C: Storing and retrieving vectors as high-dimensional points with fast nearest-neighbor lookup
- Option D: Training foundation models from scratch
Show answer and source
Answer: C. Vector databases are typically powered by k-nearest neighbor (k-NN) indexes using algorithms such as HNSW and IVF.
- A (incorrect): Translation is not a vector database function.
- B (incorrect): Foreign keys are a relational feature and do not provide similarity search.
- C (correct): Correct. Vector databases store vectors as high-dimensional points and add efficient nearest-neighbor lookup.
- D (incorrect): Vector databases store and search embeddings; they do not pre-train models.
Official sourceaws.amazon.com
Vector databases provide the ability to store and retrieve vectors as high-dimensional points.
aws.amazon.com/what-is/vector-databases/ (official page)Question 8 of 20 · D2: Fundamentals of GenAI · aif-d2-004
When Amazon Bedrock Knowledge Bases ingests documents, what does it do first so the content can be retrieved efficiently?
- Option A: Fine-tunes the foundation model on the documents
- Option B: Splits the documents into manageable chunks
- Option C: Translates every document into English
- Option D: Deletes documents older than one year
Show answer and source
Answer: B. Chunks are converted to embeddings and written to a vector index; strategies include standard, hierarchical, and semantic chunking.
- A (incorrect): Knowledge bases use retrieval; they do not fine-tune the model.
- B (correct): Correct. Bedrock first splits content into chunks, which are then converted to embeddings.
- C (incorrect): Translation is not part of ingestion.
- D (incorrect): No such retention step is part of ingestion.
Official sourcedocs.aws.amazon.com
When ingesting your data, Amazon Bedrock first splits your documents or content into manageable chunks for efficient data retrieval.
docs.aws.amazon.com/bedrock/latest/userguide/kb-chunking.html (official page)Question 9 of 20 · D2: Fundamentals of GenAI · aif-d2-005
Why can transformer-based large language models be trained much faster on GPUs than earlier recurrent neural networks (RNNs)?
- Option A: Transformers do not use neural networks.
- Option B: Transformers process entire sequences in parallel instead of one input at a time.
- Option C: Transformers need no training data.
- Option D: Transformers only work with labeled data.
Show answer and source
Answer: B. Unlike RNNs, which process inputs sequentially, transformers use self-attention over the whole sequence at once.
- A (incorrect): Transformers are neural network architectures.
- B (correct): Correct. Parallel processing of whole sequences lets GPUs speed up training.
- C (incorrect): LLMs are pre-trained on vast amounts of data.
- D (incorrect): Transformers can learn from unlabeled data through self-learning.
Official sourceaws.amazon.com
Unlike earlier recurrent neural networks (RNN) that sequentially process inputs, transformers process entire sequences in parallel.
aws.amazon.com/what-is/large-language-model/ (official page)Question 10 of 20 · D3: Applications of Foundation Models · aif-d3-001
A legal team uses an FM to draft contract clauses and wants consistent, predictable wording instead of creative variation. How should it set the temperature inference parameter?
- Option A: Use a higher temperature value
- Option B: Set the response length to the maximum
- Option C: Remove the system prompt
- Option D: Use a lower temperature value
Show answer and source
Answer: D. Temperature reshapes the probability distribution over next tokens; lower values make outputs more deterministic.
- A (incorrect): A higher value raises the chance of lower-probability tokens, increasing variation.
- B (incorrect): Response length limits output size; it does not control randomness.
- C (incorrect): Removing instructions usually makes outputs less controlled, not more consistent.
- D (correct): Correct. A lower value influences the model to select higher-probability outputs, giving more predictable text.
Official sourcedocs.aws.amazon.com
Choose a lower value to influence the model to select higher-probability outputs.
docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html (official page)Question 11 of 20 · D3: Applications of Foundation Models · aif-d3-002
Which inference parameter sets the number of most-likely candidate tokens the model considers for the next token?
- Option A: Top P
- Option B: Temperature
- Option C: Top K
- Option D: Response length
Show answer and source
Answer: C. For example, Top K = 50 makes the model select from the 50 most probable next tokens.
- A (incorrect): Top P is a percentage of cumulative probability, not a fixed count of candidates.
- B (incorrect): Temperature reshapes the probability distribution; it does not set a count.
- C (correct): Correct. Top K is the number of most-likely candidates considered for the next token.
- D (incorrect): Response length limits how many tokens are returned.
Official sourcedocs.aws.amazon.com
The number of most-likely candidates that the model considers for the next token.
docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html (official page)Question 12 of 20 · D3: Applications of Foundation Models · aif-d3-003
A model is configured with Top P = 0.8. What does this mean?
- Option A: The model returns exactly 80 tokens.
- Option B: The model samples only from the most-likely next tokens that together make up the top 80% of the cumulative probability distribution.
- Option C: The model considers exactly 8 candidate tokens.
- Option D: The model is 80% accurate.
Show answer and source
Answer: B. Lower Top P values consider a smaller share of the cumulative probability distribution.
- A (incorrect): That would be a length setting.
- B (correct): Correct. Top P limits sampling to the top P% of the cumulative probability distribution, so 0.8 means the top 80%.
- C (incorrect): A fixed count of candidates is Top K.
- D (incorrect): Top P is a sampling setting, not an accuracy measure.
Official sourcedocs.aws.amazon.com
In technical terms, the model computes the cumulative probability distribution for the set of responses and considers only the top P% of the distribution.
docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html (official page)Question 13 of 20 · D3: Applications of Foundation Models · aif-d3-004
A chatbot sometimes produces very long answers. Which type of inference parameter limits the minimum or maximum number of tokens returned?
- Option A: Response length
- Option B: Temperature
- Option C: Top K
- Option D: Embedding dimension
Show answer and source
Answer: A. Length parameters include response length, penalties, and stop sequences.
- A (correct): Correct. Response length specifies the minimum or maximum number of tokens in the generated response.
- B (incorrect): Temperature controls randomness.
- C (incorrect): Top K limits candidate tokens per step, not total length.
- D (incorrect): Embedding dimension is a property of vectors, not of generated responses.
Official sourcedocs.aws.amazon.com
An exact value to specify the minimum or maximum number of tokens to return in the generated response.
docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html (official page)Question 14 of 20 · D3: Applications of Foundation Models · aif-d3-005
What is Retrieval Augmented Generation (RAG)?
- Option A: Retraining the full model every time a document changes
- Option B: Optimizing LLM output so it references an authoritative knowledge base outside its training data before generating a response
- Option C: Compressing a large model into a smaller one
- Option D: Labeling images with bounding boxes
Show answer and source
Answer: B. RAG retrieves relevant information at query time and adds it to the prompt.
- A (incorrect): RAG avoids retraining the model.
- B (correct): Correct. This is the AWS definition of RAG.
- C (incorrect): That is model distillation.
- D (incorrect): That is data labeling.
Official sourceaws.amazon.com
references an authoritative knowledge base outside of its training data sources before generating a response.
aws.amazon.com/what-is/retrieval-augmented-generation/ (official page)Question 15 of 20 · D3: Applications of Foundation Models · aif-d3-006
A company wants its LLM application to answer questions about internal policies that change monthly, without retraining the model each time. Which approach fits best?
- Option A: Pre-training a new foundation model each month
- Option B: Raising the temperature
- Option C: Retrieval Augmented Generation (RAG) over the policy documents
- Option D: Reducing the response length
Show answer and source
Answer: C. Updating the knowledge base updates answers immediately; no model weights change.
- A (incorrect): Pre-training is the most expensive option and is unnecessary for changing documents.
- B (incorrect): Temperature does not add knowledge.
- C (correct): Correct. RAG extends an LLM to an organization's internal knowledge base without retraining.
- D (incorrect): Shorter responses do not supply new facts.
Official sourceaws.amazon.com
RAG extends the already powerful capabilities of LLMs to specific domains or an organization's internal knowledge base, all without the need to retrain the model.
aws.amazon.com/what-is/retrieval-augmented-generation/ (official page)Question 16 of 20 · D4: Guidelines for Responsible AI · aif-d4-001
AWS defines responsible AI using a set of dimensions. Which dimension means "considering impacts on different groups of stakeholders"?
- Option A: Controllability
- Option B: Privacy and security
- Option C: Veracity and robustness
- Option D: Fairness
Show answer and source
Answer: D. Other dimensions include explainability, safety, governance, and transparency.
- A (incorrect): Controllability is about mechanisms to monitor and steer AI behavior.
- B (incorrect): This dimension covers obtaining, using, and protecting data and models.
- C (incorrect): This covers correct outputs, even with unexpected or adversarial inputs.
- D (correct): Correct. AWS defines fairness as considering impacts on different groups of stakeholders.
Official sourceaws.amazon.com
Fairness Considering impacts on different groups of stakeholders
aws.amazon.com/ai/responsible-ai/ (official page)Question 17 of 20 · D4: Guidelines for Responsible AI · aif-d4-002
An operations team must be able to monitor an AI system and steer its behavior when needed. Which AWS responsible AI dimension does this describe?
- Option A: Controllability
- Option B: Fairness
- Option C: Explainability
- Option D: Safety
Show answer and source
Answer: A. Controllability gives people mechanisms to intervene in AI system behavior.
- A (correct): Correct. Controllability means having mechanisms to monitor and steer AI system behavior.
- B (incorrect): Fairness concerns impacts on different stakeholder groups.
- C (incorrect): Explainability is understanding and evaluating outputs.
- D (incorrect): Safety is preventing harmful output and misuse.
Official sourceaws.amazon.com
Controllability Having mechanisms to monitor and steer AI system behavior
aws.amazon.com/ai/responsible-ai/ (official page)Question 18 of 20 · D4: Guidelines for Responsible AI · aif-d4-003
A model must keep producing correct outputs even when it receives unexpected or adversarial inputs. Which AWS responsible AI dimension is this?
- Option A: Fairness
- Option B: Governance
- Option C: Veracity and robustness
- Option D: Controllability
Show answer and source
Answer: C. Robustness testing includes adversarial and perturbed inputs.
- A (incorrect): Fairness concerns impacts on stakeholder groups.
- B (incorrect): Governance is about incorporating best practices into the AI supply chain and processes.
- C (correct): Correct. AWS defines it as achieving correct outputs even with unexpected or adversarial inputs.
- D (incorrect): Controllability is about monitoring and steering.
Official sourceaws.amazon.com
Veracity and robustness Achieving correct system outputs, even with unexpected or adversarial inputs
aws.amazon.com/ai/responsible-ai/ (official page)Question 19 of 20 · D5: Security, Compliance, and Governance for AI Solutions · aif-d5-001
Under the AWS shared responsibility model, who is responsible for preventing prompt injection vulnerabilities in an application built on Amazon Bedrock?
- Option A: AWS
- Option B: The foundation model provider only
- Option C: No one; prompt injection cannot be mitigated
- Option D: The customer
Show answer and source
Answer: D. AWS handles security of the cloud; customers handle security in the cloud, including their application.
- A (incorrect): AWS secures the underlying infrastructure and the Bedrock service, not your application code.
- B (incorrect): The model provider does not write your application.
- C (incorrect): Input validation, secure coding, and testing mitigate it.
- D (correct): Correct. AWS states that preventing vulnerabilities like prompt injection lies with the customer.
Official sourcedocs.aws.amazon.com
the responsibility for secure application development and preventing vulnerabilities like prompt injection lies with the customer.
docs.aws.amazon.com/bedrock/latest/userguide/prompt-injection.html (official page)Question 20 of 20 · D5: Security, Compliance, and Governance for AI Solutions · aif-d5-002
A company wants applications in its VPC to call Amazon Bedrock privately, without an internet gateway, NAT device, or public IP addresses. What should it use?
- Option A: AWS Artifact
- Option B: Interface VPC endpoints powered by AWS PrivateLink
- Option C: Amazon Macie
- Option D: A public S3 bucket
Show answer and source
Answer: B. You create an interface endpoint; instances do not need public IP addresses.
- A (incorrect): Artifact provides compliance documents.
- B (correct): Correct. PrivateLink creates a private connection between your VPC and Bedrock.
- C (incorrect): Macie discovers sensitive data in S3.
- D (incorrect): A public bucket increases exposure and does not provide private connectivity.
Official sourcedocs.aws.amazon.com
You can access Amazon Bedrock as if it were in your VPC, without the use of an internet gateway, NAT device, VPN connection
docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (official page)
After the sampler
The full AIF-C01 bank has 260 questions and four timed 65-question mocks. Paid access is coming soon; see pricing. Found a problem in a question? Use “Report a problem” in the practice tool or email busywright@mail.grokbot.com. Fixes are logged on the errata page.
DocBacked is unofficial and is not affiliated with, or endorsed by, the exam provider. These are original practice questions, not real exam questions.