How to use this post
Eight original AWS Certified AI Practitioner (AIF-C01) scenario questions. Answer each one before reading its explanation. Every explanation covers why the right answer is right and why each wrong answer is wrong.
The AIF-C01 is a foundational exam, but it isn't a trivia quiz. It rewards people who can reason about when to use a foundation model rather than a traditional ML model, how to reduce hallucinations, what responsible AI means in an AWS context, and which service fits a given job.
These eight questions are written to practise that reasoning. They are not real exam questions. Using leaked exam content breaks the AWS Certification Program Agreement and can cost you your certification, and memorised answers don't survive a reworded question anyway.
AIF-C01 at a glance
- Questions: 65 (50 scored, 15 unscored)
- Time: 90 minutes
- Passing score: 700 on a scaled 100–1,000 range. It isn't a percentage.
- Cost: $100 USD
- Level: Foundational, with no prerequisites
- Domains: (1) Fundamentals of AI and ML, (2) Fundamentals of Generative AI, (3) Applications of Foundation Models, (4) Guidelines for Responsible AI, (5) Security, Compliance, and Governance for AI Solutions
Check the official AWS Certified AI Practitioner page close to your exam date. AWS updates the details from time to time.
The 8 questions
Question 1: AI vs. ML vs. deep learning
A company wants to categorise incoming support emails as "billing", "technical" or "general". It has 40,000 past emails, each already labelled with the correct category. Which approach best describes what it should build?
- A. A rules engine with hand-written keyword lists for each category
- B. A supervised machine learning classification model trained on the labelled emails
- C. An unsupervised clustering model to discover categories automatically
- D. A reinforcement learning agent that learns from user rewards
Correct answer: B
The giveaway is already labelled with the correct category. Labelled historical data plus a fixed set of output classes is the textbook definition of a supervised classification problem.
Why A is wrong: A rules engine is brittle and isn't machine learning. It can't generalise to phrasings its authors didn't anticipate, and it wastes the labelled data.
Why C is wrong: Clustering is for when you don't have labels and want to discover groupings. Here the categories are already known.
Why D is wrong: Reinforcement learning is for sequential decisions with a reward signal, such as games or robotics, not one-shot text classification.
Question 2: When not to use generative AI
A bank needs a system to approve or deny loan applications. The decision must be explainable to regulators, reproducible, and based on a fixed set of financial rules. Which is the most appropriate choice?
- A. A large language model prompted with each applicant's details
- B. A generative AI chatbot fine-tuned on past decisions
- C. A traditional, interpretable ML model (or rules-based system) with documented features
- D. An image-generation model
Correct answer: C
This tests whether you know the limits of generative AI. Explainable, reproducible, rule-based and regulator-facing are exactly the conditions where an LLM is a poor fit: it can give different outputs for the same input, is hard to fully explain, and can hallucinate. An interpretable model or a documented rules engine gives auditable, consistent decisions.
Why A is wrong: It relies on an LLM for a high-stakes, regulated, deterministic decision.
Why B is wrong: Fine-tuning doesn't make a generative model reproducible or explainable.
Why D is wrong: It has nothing to do with the task. The exam often rewards the answer that says generative AI is the wrong tool.
Question 3: Reducing hallucinations with your own data
A company runs a chatbot on Amazon Bedrock to answer employee questions about internal HR policies. The model keeps inventing policies that don't exist. The company wants answers grounded in its actual policy documents, without retraining the model. What should it implement?
- A. Increase the model's temperature setting
- B. Retrieval Augmented Generation (RAG) using the policy documents as a knowledge source
- C. Switch to a larger foundation model
- D. Add more few-shot examples of a polite tone to the prompt
Correct answer: B
RAG grounds answers in specific, trusted documents without retraining. Relevant passages are retrieved from the policy documents and given to the model as context, so answers come from real content rather than the model's memory. On AWS, Amazon Bedrock Knowledge Bases is built for this.
Why A is wrong: A higher temperature makes output more random, which makes hallucination worse.
Why C is wrong: A bigger model still has no access to the company's private policies, so it can still invent them.
Why D is wrong: Tone examples change style, not factual grounding.
Question 4: Prompt engineering technique
A developer wants a foundation model to solve multi-step word problems more accurately by having it show its reasoning step by step before giving the final answer. Which prompt engineering technique is this?
- A. Zero-shot prompting
- B. Chain-of-thought prompting
- C. Prompt injection
- D. Temperature tuning
Correct answer: B
Asking the model to work through its reasoning step by step before answering is chain-of-thought prompting. It improves performance on multi-step reasoning such as maths and logic.
Why A is wrong: Zero-shot means giving the task with no examples. It doesn't describe the show-your-steps behaviour.
Why C is wrong: Prompt injection is an attack in which malicious input overrides the intended instructions. Watch for a security term used as bait.
Why D is wrong: Temperature adjusts randomness, not reasoning structure.
Question 5: Choosing the right AWS AI service
A retailer wants to extract text from scanned paper invoices and pull out fields such as vendor name, date and total. It wants a managed service and needs minimal ML expertise. Which AWS service fits best?
- A. Amazon Textract
- B. Amazon Comprehend
- C. Amazon Polly
- D. Amazon Rekognition
Correct answer: A
Amazon Textract extracts text, handwriting and structured data (forms, tables, key-value pairs) from scanned documents: exactly the invoice use case.
Why B is wrong: Comprehend analyses text that already exists (sentiment, entities, key phrases). It doesn't pull text off a scanned image.
Why C is wrong: Polly turns text into speech, the wrong direction entirely.
Why D is wrong: Rekognition analyses images and video and can detect text in images, but for document and form field extraction Textract is the purpose-built service.
Question 6: Responsible AI: identifying bias
A hiring model was trained mostly on résumés from one demographic group and now scores candidates from other groups lower, even when their qualifications are equal. Which responsible-AI concern does this most directly illustrate?
- A. Model overfitting to the test set
- B. Bias and fairness arising from non-representative training data
- C. High inference latency
- D. Insufficient prompt length
Correct answer: B
Training data that over-represents one group and produces systematically worse outcomes for others is a classic bias and fairness problem. The fixes include more representative data, fairness evaluation and ongoing monitoring. Amazon SageMaker Clarify helps detect this kind of bias.
Why A is wrong: Overfitting is a general accuracy problem. The scenario describes a demographic disparity, which is a fairness issue.
Why C is wrong: Latency is a performance concern, unrelated to the harm described.
Why D is wrong: Prompt length has nothing to do with a trained model's bias.
Question 7: The customisation ladder
A team using a foundation model on Amazon Bedrock finds that prompt engineering alone can't produce the specialised legal tone and terminology it needs, and RAG doesn't fully solve it either. It has a large, high-quality dataset of domain examples and wants the model itself to internalise the style. Which approach is this?
- A. Increasing the context window
- B. Fine-tuning the model on the domain dataset
- C. Lowering the temperature to zero
- D. Switching to a smaller model
Correct answer: B
When prompting and RAG aren't enough and you have a substantial, high-quality dataset, fine-tuning adjusts the model's weights so it internalises the domain's style and terminology. It's the heavier, more expensive option, which is why the exam expects you to reach for it only after lighter options fall short: prompt engineering, then RAG, then fine-tuning.
Why A is wrong: A bigger context window lets you pass more text, but doesn't teach the model a style.
Why C is wrong: Temperature zero makes output more deterministic, not more expert.
Why D is wrong: A smaller model generally reduces capability.
Question 8: Security and governance
A company wants to be sure that the prompts and completions it sends to a foundation model on Amazon Bedrock aren't used to train the base model. Which statement reflects how Amazon Bedrock handles this?
- A. All data sent to any foundation model is automatically shared publicly
- B. Amazon Bedrock doesn't use prompts and completions to train the underlying base models, and customers keep control of their data
- C. The only way to keep data private is to build your own model from scratch
- D. Prompt data must be emailed to AWS Support for review
Correct answer: B
A key governance point for AIF-C01: Amazon Bedrock doesn't use your prompts and completions to train the underlying foundation models, and your data stays under your control, protected by AWS security and privacy controls.
Why A is wrong: It contradicts AWS's stated data handling.
Why C is wrong: An overreaction. You don't need your own model to keep data private.
Why D is wrong: A deliberately absurd distractor: a free point if you know the real policy.
How to use questions like these
- Answer before you read the explanation. Committing to an answer is what builds recall.
- For every wrong option, ask why it's wrong. The exam is built on plausible distractors, and eliminating them is half the skill.
- Turn each question into a concept. Question 3 isn't about one chatbot. It's about RAG grounding answers in your own data without retraining.
- Learn a one-line purpose for each AWS AI service: Textract, Comprehend, Rekognition, Polly, Translate, Transcribe, Bedrock and SageMaker.
- Do a full-length timed set near the end of your preparation, so 65 questions in 90 minutes feels routine.
Try 20 more, free
→ Take the free AIF-C01 practice test: 20 questions, no signup, and your score as soon as you submit. Give your email at the end to see every answer explained.
Full AIF-C01 practice tests
333 questions across 5 full-length practice tests, each with an explanation for every answer. See the AIF-C01 course →
This article is not affiliated with or endorsed by Amazon Web Services. All questions above are original.