Try 30 sample questions
These are real samples from the AIF-C01 bank — the full version has 865 like these.
1. What is machine learning?
- Software written with hand-coded rules for every possible input
- Computer systems that improve their performance on a task by learning patterns from data
- Software that only runs on specialized AI chips like GPUs
- Any program that uses a neural network
Correct answer: B. Machine learning systems learn their behavior from data rather than executing fixed instructions written by developers, so their performance improves as they see more examples. Why others are wrong: A) Hand-coded rules are the opposite of machine learning; they describe traditional programming, where humans write every instruction.; C) Machine learning is a software approach, not a hardware requirement; models can run on CPUs and many kinds of hardware.; D) Neural networks are one machine learning technique (deep learning), but ML includes many methods that do not use neural networks.
2. A team wants to build a model that reads handwritten digits, but it has only unlabeled images of handwriting. What is TRUE?
- It can train a supervised classifier immediately
- Supervised learning works fine without labels
- Labels are only needed for deep learning
- It must first obtain labeled examples (or label the images) before supervised training
Correct answer: D. Supervised learning is defined by training on labeled data. Without labels indicating which digit each image shows, the team must create or obtain them first. Why others are wrong: A) A supervised classifier cannot be trained without labels to learn from.; B) Labels are what make learning 'supervised'; without them it is a different learning type.; C) All supervised methods need labels, not just deep learning.
3. A company has millions of raw product reviews with no sentiment ratings attached. Which term describes this data?
- Data that has been held out from training for final evaluation
- Data in which the features have not been extracted yet
- Data that has been anonymized for privacy
- Data consisting of examples with no known correct answer attached
Correct answer: D. Unlabeled data consists of examples without known answers, such as raw customer reviews with no sentiment rating. It is the typical input for unsupervised learning, where the model finds structure like clusters on its own. Why others are wrong: A) Held-out evaluation data is the test dataset. Unlabeled data is defined by the absence of known correct answers, not by being set aside for evaluation.; B) Feature extraction creates the input variables, which is a separate step. Data without known correct answers attached is unlabeled, whether or not features have been extracted.; C) Anonymization removes identifying details but examples can still carry labels; unlabeled means no known answers, not no identities.
4. A fraud model has 95% precision but only 30% recall. What does this mean in practice?
- The model finds most positives but its positive predictions are often wrong
- The model is perfectly calibrated
- The model predicts every case as positive
- The model's positive predictions are usually correct, but it misses many actual positive cases
Correct answer: D. High precision with low recall means the model's positive predictions are trustworthy, but it finds only a small fraction of the real positives - it is conservative and misses many cases. Why others are wrong: A) That describes high recall with low precision - the reverse of this model's behavior.; B) Calibration concerns whether predicted probabilities match observed frequencies; precision and recall describe different properties.; C) Predicting everything as positive would produce very low precision, not 95%.
5. An enterprise has documentation spread across wikis, PDFs, and internal sites. Employees need to ask questions in natural language and get answers with links to the source documents. Which AWS service provides intelligent enterprise search?
- Amazon Comprehend
- Amazon Personalize
- Amazon Kendra
- Amazon Textract
Correct answer: C. Amazon Kendra is an intelligent search service that uses natural language processing to understand questions and return answers from enterprise content, with links to the source documents. Why others are wrong: A) Amazon Comprehend analyzes text but does not index an organization's documents for search.; B) Amazon Personalize recommends items to users; it does not search enterprise documentation.; D) Amazon Textract extracts text from scanned documents but provides no question-answering search experience.
6. To train a model that detects manufacturing defects from photos of products on an assembly line, what training data is needed?
- Spreadsheets of monthly sales figures
- Audio recordings of the factory floor
- Labeled images showing defective and non-defective products
- Customer survey text responses
Correct answer: C. An image-based defect detector must learn from images, and supervised training requires those images to be labeled as defective or non-defective. Why others are wrong: A) Sales spreadsheets contain no visual information about product defects.; B) Audio recordings contain no images of the products for the model to learn from.; D) Survey text describes opinions, not the visual appearance of defects.
7. Which statement BEST distinguishes pre-training from fine-tuning?
- Pre-training adapts a model to one task, while fine-tuning trains it on broad data
- They are two names for the same process
- Fine-tuning always requires more data and compute than pre-training
- Pre-training teaches general patterns from large, broad data, while fine-tuning adapts the model to a specific task
Correct answer: D. Pre-training builds the foundation: general capabilities learned from massive datasets. Fine-tuning takes that foundation and specializes it with smaller, task-specific datasets. Why others are wrong: A) This reverses the two. Pre-training is the broad phase, and fine-tuning is the specializing phase.; B) They are distinct phases with different goals, data, and cost profiles.; C) The reverse is true. Pre-training is the vastly more expensive phase, while fine-tuning uses far less data and compute.
8. Which statement BEST distinguishes diffusion models from large language models?
- Diffusion models are primarily associated with generating images, while LLMs are primarily associated with generating text
- Diffusion models are trained only on video data
- Diffusion models understand language better than large language models
- Diffusion models do not use neural networks
Correct answer: A. The practical distinction at awareness level is modality focus: diffusion architectures excel at image generation, often guided by text prompts, while LLMs specialize in language. Why others are wrong: B) Diffusion models are most commonly trained on still images paired with text; the distinction in the question is about what each model type generates.; C) Diffusion models use text encoders to follow prompts, but language understanding and generation remain the LLM's primary strength.; D) Diffusion models are neural networks trained with deep learning, just like LLMs.
9. A generative AI model confidently states a false fact in fluent, convincing language. What is this phenomenon called?
- Hallucination
- Overfitting
- Underfitting
- Regularization
Correct answer: A. Hallucination is the term for a generative model producing confident-sounding but false or fabricated content, such as invented facts or citations. Why others are wrong: B) Overfitting means a model memorized training data too closely and performs poorly on new data; it is not about fluent falsehoods.; C) Underfitting means a model is too simple to capture patterns in the data; it is unrelated to confident false statements.; D) Regularization is a training technique to reduce overfitting, not a name for false outputs.
10. A company wants to BUILD its own custom generative AI application on top of foundation models, rather than deploy a ready-made assistant. Which AWS service is the correct starting point?
- Amazon Q Business
- Amazon Bedrock
- Amazon SageMaker, by training a custom model
- PartyRock
Correct answer: B. Amazon Bedrock provides API access to foundation models from multiple providers, which is what developers use to build their own generative AI applications, features, and workflows. Why others are wrong: A) Amazon Q Business is a ready-made AI assistant for enterprise knowledge work. The company wants to build a custom application, not deploy a packaged assistant.; C) Training a custom model is a different project; the company wants to build on existing foundation models, which is what Bedrock provides.; D) PartyRock is a no-code playground for experimenting with small generative AI apps. It is not the production API platform for building a company's custom application.
11. Which order best describes the stages of a generative AI project?
- Define the use case, select and adapt a model, evaluate outputs, deploy, then monitor
- Deploy first, then define the use case
- Monitor the system before any use case exists
- Evaluate outputs before choosing a model or use case
Correct answer: A. GenAI projects flow from problem definition to model selection and adaptation, then evaluation, deployment, and ongoing monitoring - each stage depends on the previous one. Why others are wrong: B) Deploying before defining the problem means building a solution with no purpose.; C) There is nothing to monitor before a system exists.; D) Evaluation requires a model and a defined task to evaluate against.
12. A company has 10,000 product manuals and wants a pilot where technicians type a question in their own words and get the relevant manual section. Which approach makes this work?
- Convert the manuals to embeddings and retrieve sections by meaning similarity to the question
- Build a keyword search index over the manual text
- Train a new foundation model from scratch on the manuals
- Have technicians browse a clickable table of contents to find sections themselves
Correct answer: A. Embeddings capture the meaning of text as vectors, so a semantic search can find the manual section closest in meaning to the question even when the words differ. Why others are wrong: B) Keyword search matches words, not meaning, so a technician's paraphrased question would miss sections that use different wording.; C) Training a foundation model from scratch is enormously expensive and unnecessary when retrieval over embeddings solves the task.; D) Manual browsing does not answer a typed question automatically; the pilot requires the system to return the relevant section.
13. A refund workflow automatically reads a model's one-word answers ('approved' or 'denied'), but the model sometimes adds extra explanations that break the parser. What is the BEST fix?
- Increase the temperature for more creative answers
- Ask the model to write longer explanations
- Switch to a larger, more capable model and trust it to follow the format
- Instruct the model to output only the single word, use a low temperature, and validate the output format programmatically
Correct answer: D. Clear format instructions plus low temperature make outputs more deterministic, and programmatic validation catches any deviation before it breaks the workflow. Why others are wrong: A) Higher temperature increases variability, making format violations more likely.; B) Longer explanations are exactly what breaks the parser.; C) Even strong models occasionally add extra text; programmatic validation is what makes the workflow robust.
14. A retailer is adding an AI chat assistant to its checkout page, where shoppers expect an answer in under two seconds. Which model selection criterion is most important here?
- The model's ability to generate very long essays
- The model's training data cutoff date from three years ago
- The model's parameter count, preferring larger models for better checkout answers
- Low response latency, even if it means choosing a smaller, faster model
Correct answer: D. A checkout chat has a strict latency requirement; a smaller, faster model that still answers well is preferable to a slower, more capable model that makes shoppers wait. Why others are wrong: A) Long-form essay generation is unrelated to quick checkout answers and tends to increase latency.; B) The training cutoff does not determine response speed; latency comes from model size and serving setup.; C) Preferring larger models for quality is reasonable in general, but here parameter count directly conflicts with the two-second latency requirement.
15. A user asks about 'ways to lower my electricity bill' and the system retrieves a document titled 'home energy saving tips' even though it shares no exact keywords with the query. What capability made this match?
- Keyword counting
- Sorting the documents by title before searching
- Searching the documents with keyword matching
- Semantic search with embeddings, matching meaning rather than exact words
Correct answer: D. Embeddings capture meaning as vectors, so a query about lowering electricity bills is close in vector space to energy-saving content even without shared keywords. Why others are wrong: A) Keyword counting requires shared terms; here the match succeeded despite different wording.; B) Sorting by title only reorders results; it says nothing about which document answers the query.; C) Keyword matching needs shared terms; this match succeeded despite different wording, which is semantic search.
16. A team uses Amazon Bedrock and wants to adapt a supported foundation model using its own labeled examples of ideal customer-service responses. Which Bedrock customization option should it use?
- Continued pre-training with unlabeled data
- Fine-tuning with the labeled examples
- Changing the model's token limit
- Continued pre-training on the support conversation transcripts
Correct answer: B. Amazon Bedrock supports fine-tuning for certain models, where you provide labeled examples and AWS trains a customized copy of the model. Labeled ideal responses are exactly the input fine-tuning expects. Why others are wrong: A) Continued pre-training on Bedrock uses unlabeled text for domain adaptation, not labeled examples of desired behavior.; C) The token limit is a model property, not a customization option, and changing it would not teach customer-service behavior.; D) Continued pre-training on Bedrock uses unlabeled text for domain adaptation; it does not take labeled examples of desired behavior the way fine-tuning does.
17. What is a key benefit of breaking a complex task into a chain of smaller prompts instead of using one giant prompt?
- It reduces the number of calls the application makes to the model
- Once a step runs, its output no longer needs checking
- It removes the need to validate the final output
- Each step's output can be checked, corrected, and fed into the next step
Correct answer: D. Prompt chaining makes each stage's output visible, so teams can validate, correct, or retry individual steps before passing results along. This control and debuggability is the main advantage over a single monolithic prompt. Why others are wrong: A) Chaining splits one call into several, so it increases the number of model calls; control and debuggability are the benefits.; B) Each step can still produce errors; chaining lets teams check and correct outputs, but it does not make them automatically correct.; C) Every step's output, including the final one, may need checking; chaining improves visibility but does not remove the need for validation.
18. A company uses a well-known safe foundation model in its customer service application. Why might it STILL add guardrails to the application?
- To make the model run faster
- To enforce the company's own policies and compliance rules, such as blocking specific topics the business does not want discussed
- To reduce the number of tokens in every prompt
- Demonstrate to regulators that human reviewers are no longer required
Correct answer: B. Model-level safety is general; applications add guardrails to enforce their specific policies, brand rules, and regulatory requirements, such as denying particular topics. Why others are wrong: A) Guardrails add safety checks; they do not increase inference speed.; C) Guardrails do not shorten prompts; prompt size is managed through prompt engineering.; D) Guardrails enforce company policies but complement, not replace, human oversight; high-stakes uses still need review.
19. When evaluating two foundation models for a customer support chatbot, which measure BEST reflects business value?
- The models' parameter counts
- Each model's score on general language benchmarks
- The percentage of customer issues resolved without human escalation
- The average response time of each model
Correct answer: C. Resolution rate without escalation directly measures the chatbot's business purpose: solving customer problems, making it the best business-value indicator. Why others are wrong: A) Parameter counts describe model size, not how well the chatbot serves customers.; B) Benchmark scores describe general capability, not whether customer issues get resolved; resolution rate ties directly to business value.; D) Speed is an operational detail; the business cares whether issues are resolved without costly human escalation.
20. A product-description generator feeds another application that parses its output as JSON. Free-form paragraphs keep breaking the pipeline. Which approach BEST fixes this?
- Switch to a larger foundation model
- Increase the temperature for more output variety
- Show the model many examples of correct JSON, since larger training exposure makes formatting errors disappear on its own
- Instruct the model to return JSON matching the required schema, and have the application validate the JSON before accepting it
Correct answer: D. Machine consumers need machine-readable contracts: explicit structured-output instructions make the model emit JSON in the required shape, and application-side validation catches and rejects any malformed response instead of letting it break downstream processing. Why others are wrong: A) A larger model still writes prose unless instructed otherwise; the failure is about format, not capability.; B) More randomness makes the output format less predictable, worsening the parsing failures.; C) Examples improve formatting odds but give no guarantee; only explicit schema instructions plus application-side validation make the pipeline reliable.
21. A logistics company uses RAG to summarize long carrier contracts. The legal team cares that each summary preserves the contract's meaning, not that it copies exact wording. Which evaluation approach best fits?
- Count how many words overlap between the summary and the contract (ROUGE only)
- Use BERTScore and human review, since meaning and correctness matter more than word overlap
- Use ROUGE and BERTScore only, with no human review
- Use BLEU scores, which are designed for machine translation
Correct answer: B. BERTScore measures semantic similarity rather than surface overlap, and human review confirms legal correctness, which automated overlap scores cannot guarantee. Why others are wrong: A) Word overlap can miss a summary that is worded differently but legally wrong, or correct but rephrased.; C) Automated scores alone cannot confirm legal correctness; the legal team needs human review to verify the summary truly preserved the contract's meaning.; D) BLEU was designed for translation quality, not for judging contract summaries.
22. A bank builds a loan model that predicts 'creditworthiness' using a borrower's zip code as a stand-in, because it lacks direct repayment data. What is the problem with this approach?
- Zip codes are objective, so the model using them is fair
- The model uses too few features to be accurate
- The bank collected the zip codes without consent
- Measurement bias: the zip code is a flawed proxy that can stand in for protected characteristics like race
Correct answer: D. When the thing measured is only a proxy for the real target, the proxy can carry hidden correlations with protected attributes, baking unfairness into the model. Why others are wrong: A) An objective-looking input can still act as a proxy for a protected characteristic.; B) The problem is not feature count; it is that the zip code stands in for protected characteristics.; C) Consent is a separate issue; the fairness problem is the flawed proxy measurement.
23. A company publishes a model card for its customer-support chatbot. Which information should the model card include?
- The model's per-user inference cost and billing rates
- The model's intended use, training data sources, evaluation results, and known limitations
- A sample of raw customer prompts taken from production
- The hyperparameter values used in the final training run
Correct answer: B. Model cards are transparency artifacts that document what a model is for, what data trained it, how it performed in evaluation, and where it falls short, so deployers can judge whether it fits their use case. Why others are wrong: A) Pricing describes the commercial offering, not the model; a model card documents intended use, data, evaluation, and limitations.; C) Sharing real customer prompts would leak private conversations; a model card summarizes data sources without exposing individual records.; D) Hyperparameters are engineering detail; a model card focuses on what the model is for, what it was trained on, how it performed, and its limits.
24. A company building a customer-service chatbot decides to train it only on anonymized support transcripts and to avoid collecting customers' full names and addresses. Which privacy principle is this?
- Data maximization
- Data monetization
- Data minimization: collecting only the personal data truly needed for the purpose
- Purpose limitation, because the data will only be used for customer service
Correct answer: C. Data minimization means limiting personal data collection to what the purpose requires; training on anonymized transcripts without unnecessary identifiers reduces privacy risk if data is ever exposed. Why others are wrong: A) Data maximization is the opposite - collecting as much as possible.; B) Monetization concerns selling data, not limiting its collection.; D) Purpose limitation governs what data may be used for; the described action - collecting only the data truly needed - is data minimization.
25. A credit-scoring model denies an applicant, who asks why the decision was made. Which response best reflects the explainability dimension of responsible AI?
- Refuse to say anything because the model is proprietary
- Explain that the decision was based on a complex neural network
- Provide a plain-language explanation of the main factors that influenced the decision
- Provide the applicant with the model's overall accuracy statistics
Correct answer: C. Explainability means affected people receive an understandable account of the key factors behind a decision, supporting transparency and the ability to contest it. Why others are wrong: A) Total refusal to explain undermines transparency, even when some model details remain proprietary.; B) Naming the technique does not tell the applicant which factors influenced their specific decision.; D) Aggregate accuracy says nothing about why this particular applicant was denied.
26. A team fine-tunes a model on Amazon SageMaker using sensitive customer records. Under the shared responsibility model, who is responsible for ensuring that training data is encrypted and that only authorized team members can access it?
- AWS, because it owns the SageMaker service
- The model provider, because it published the base model
- The compliance auditor, because auditors verify encryption controls
- The customer, because protecting its data and controlling access are customer responsibilities
Correct answer: D. Data protection and access control are always customer responsibilities: the customer must encrypt its training data and restrict access, even on a managed service like SageMaker. Why others are wrong: A) AWS secures the SageMaker infrastructure, not the customer's data classification or access decisions.; B) The base-model provider has no role in how the customer stores or protects its own training data.; C) Auditors verify controls after the fact; implementing encryption and access control for the training data is the customer's job.
27. Which combination best protects training data stored in Amazon S3?
- SSE-S3 encryption plus open read access for the data science team
- Individual IAM users for each team member, each with AmazonS3FullAccess
- Encryption at rest with KMS, restrictive bucket policies, and Macie scans for sensitive data
- Separate buckets per project with versioning enabled but no encryption
Correct answer: C. Defense in depth works best: KMS encryption protects data at rest, tight bucket policies limit who can reach it, and Macie watches for sensitive data that should not be there. Why others are wrong: A) Encryption at rest does not replace authorization control; open read access makes the data readable by anyone who can reach the bucket.; B) Individual users fix accountability, but S3FullAccess over-grants every bucket in the account; least privilege needs policies scoped to the dataset buckets.; D) Separation and versioning help organization and recovery, but the stored bits stay unencrypted - there is no confidentiality protection.
28. Amazon SageMaker Model Monitor can alert a team when which condition occurs?
- A new foundation model is released
- Incoming inference data drifts significantly from the training data baseline
- The model's accuracy drops below a set threshold
- Training job failures in the pipeline
Correct answer: B. Model Monitor detects data drift, concept drift, and quality degradation by comparing production traffic against baselines, alerting teams before bad predictions accumulate. Why others are wrong: A) New model releases are announced by AWS, not detected by Model Monitor.; C) Accuracy is measured against labeled ground truth in quality monitoring; Model Monitor's signature alert is data drifting from the training baseline.; D) Job failures are pipeline events; Model Monitor watches deployed endpoint data for drift.
29. A company with strict data-residency rules self-hosts an open-weights model on Amazon SageMaker inside its own VPC instead of calling a managed model API. Which security benefit does this provide?
- Eliminates the need to encrypt data in transit, since traffic stays inside the company's network
- AWS becomes responsible for the company's data governance
- The open-weights license transfers all compliance responsibility to the model publisher
- Prompts and data stay within the company's own VPC and chosen Region, under its own controls, instead of leaving for an external service endpoint
Correct answer: D. Self-hosting trades managed convenience for control: inference happens on infrastructure inside the customer's VPC and Region, so sensitive prompts never travel to a third-party endpoint - the core data-residency advantage. Why others are wrong: A) Internal traffic still requires encryption; self-hosting changes where data lives, not the need to protect it.; B) Under the shared responsibility model the customer always owns its data governance, whichever hosting model it chooses.; C) Self-hosting keeps data-governance and compliance responsibility with the company, not the publisher.
30. A hospital is launching an AI assistant that summarizes patient discharge notes for clinicians. Which launch plan best combines the security, compliance, and governance controls this workload requires?
- Launch with encryption and logging enabled, and add Guardrails after the pilot phase
- Encrypt the data but skip access controls, logging, and documentation to move faster
- Have the assistant flag only low-confidence summaries for clinician review
- Keep all data in the required Region with KMS encryption, enforce least-privilege IAM, apply Guardrails with PII filters, require clinician review of summaries, enable CloudTrail logging, and document the release with a Model Card and approval records
Correct answer: D. Clinical AI demands the full control stack: residency and encryption for data protection, least privilege and Guardrails for runtime safety, human review for clinical accountability, logging for audit, and Model Cards plus approvals for governance. Why others are wrong: A) Patient data requires the full control set before go-live; Guardrails cannot wait for a later phase.; B) Encryption alone without access control, logging, or documentation fails audit, accountability, and breach-investigation requirements.; C) Confidence scores do not reliably catch confident errors; patient summaries need consistent clinician review. AIF-C01 - Real-exam format questions (Q801-865) 65 additional questions in the other formats used by the real exam: multiple response (Select TWO), ordering, matching, and case-study sets. Each question includes the correct answer and full explanations. Multiple Response Questions (Select TWO)
Exam facts: AWS Certified AI Practitioner (AIF-C01)
- Exam: AWS Certified AI Practitioner (AIF-C01)
- Cost: $100 USD
- Duration: 90 minutes
- Questions: 65 — multiple choice, multiple response, ordering, and matching (50 scored, 15 unscored)
- Passing score: 700 (scaled score of 100–1,000)
- Validity: 3 years
Frequently asked questions
How long do I get access?
Three months (90 days) from purchase, covering both the interactive online bank and the full PDF download. Access can be extended.
Is there a free trial?
Yes — a free 30-question sample quiz and a free sample PDF are available with no account required, so you can judge the quality before paying.
Are these questions like the real AIF-C01 exam?
All 865 questions are mapped to the official AIF-C01 exam objectives, written in the same style and difficulty as the real exam.
Can I study offline?
Yes. Your purchase includes the complete question bank as a downloadable PDF for offline study.
What is the refund policy?
Full access is refundable within 15 days, or before 25% of the questions have been attempted — whichever comes first. Downloaded PDFs are non-refundable.
How much does the AWS AI Practitioner exam cost?
The AIF-C01 exam costs $100 USD.
How many questions are on the AIF-C01 exam?
65 questions in 90 minutes. Alongside multiple choice and multiple response, this exam can include ordering and matching questions. 50 questions are scored and 15 are unscored trial questions.
What is the passing score for AIF-C01?
700 on AWS's scaled score range of 100 to 1,000. Scoring is compensatory — you need to pass the exam overall, not each section.
How long is the AWS AI Practitioner certification valid?
Three years. Before it expires, you recertify by passing the current exam or a higher-level AWS exam.
Where can I practice for the AIF-C01 exam for free?
Right here on CertCrack. Start with the free 30-question sample quiz and the free sample PDF on this page — no account needed — and read the free AI Practitioner study guidebook. When you're ready for the full bank, this page gives you every practice question with detailed explanations for $10.