Explore / AWS Certified AI Practitioner (AIF-C01) Exam Prep / Practice questions
AWS Certified AI Practitioner (AIF-C01) practice questions
10 questions from the bank behind this course's timed mock. Free to read and to answer, no account needed, and each has its answer and an explanation. They belong to AWS Certified AI Practitioner (AIF-C01) Exam Prep.
Answer them one at a time →Read lesson 1 free
Question 1
A data analyst at an insurance company reviews a claims-prediction model. The model is almost perfectly accurate on the data it was trained on, but its accuracy drops sharply on new claims submitted this month. Which term best describes this problem?
- Inferencing
- Overfitting
- Fairness
- Underfitting
Show the answer
Answer: B. Overfitting
An overfit model memorises its training data and performs poorly on new data, which matches the gap described. Underfitting is wrong because an underfit model is too simple to capture the pattern at all, so it would not score well on the training data either.
Question 2
A bank's data team exports a table of 200,000 past loans. Each row has columns for applicant income, loan amount and term, plus a column recording whether the loan was repaid or defaulted. The team wants to describe this dataset accurately before model training. Which classification is correct?
- Labeled, unstructured text data
- Labeled, structured tabular data
- Unlabeled time-series data
- Unlabeled, structured tabular data
Show the answer
Answer: B. Labeled, structured tabular data
The rows and columns make the data structured and tabular, and the repaid/defaulted column is the known answer, so it is labeled. It is not unlabeled tabular data, because that would have no outcome column.
Question 3
A company is building an email filter where "positive" means spam. Users complain strongly when a legitimate customer email is sent to the spam folder, but they tolerate the occasional spam message reaching the inbox. Which model metric should the data scientist prioritise?
- Return on investment (ROI)
- Precision
- Accuracy
- Recall
Show the answer
Answer: B. Precision
Precision measures how many of the items flagged positive really are positive, so it is the metric to prioritise when false positives are costly. Recall guards against false negatives (missed spam), which the users here tolerate.
Question 4
A product manager at an online retailer reviews an AI assistant that drafts replies to customer emails. The replies are fluent but generic and never mention the customer's subscription tier or the relevant product documentation. The team has already reworded the instruction several times with no improvement. Which TWO changes are context engineering actions that address the root cause? (Select TWO.)
More than one answer is right.
- Retrieve relevant product documentation from a knowledge base and include it in the request
- Rewrite the instruction to use a more formal tone and a different role
- Ask the model to produce five bullet points instead of three
- Replace the foundation model with a larger one that has more parameters
- Add the customer's profile details, such as subscription level, to each request
Show the answer
Answers: A. Retrieve relevant product documentation from a knowledge base and include it in the request; E. Add the customer's profile details, such as subscription level, to each request
The output lacks information, so the fix is to supply the user profile and retrieved knowledge base documents, which are both components of a well-engineered context. Rewording the instruction is prompt engineering and cannot add facts the model was never shown.
Question 5
A wholesale distributor deploys a GenAI assistant for its sales team. A sales representative asks for the unit price of a product, and the assistant returns the price from the previous year's catalogue. That price was genuine at the time, but the foundation model was trained before the current price list was published and has no access to company data. Which disadvantage of GenAI does this BEST describe?
- Hallucinations
- Interpretability
- Inaccuracy
- Nondeterminism
Show the answer
Answer: C. Inaccuracy
The output is wrong because the training data is stale and lacks the company's current facts, which is inaccuracy. It is not a hallucination, because the price was not fabricated: it was a real figure that is now out of date.
Question 6
A legal services firm wants a foundation model to answer questions about individual contracts that are each up to 400 pages of plain English text. Pilot requests fail because the full contract cannot be submitted in one prompt. Which selection criterion should the firm re-evaluate first?
- Modality
- Support for fine-tuning
- Input/output length (context window)
- Multi-lingual support
Show the answer
Answer: C. Input/output length (context window)
The context window must hold the longest input, so a model with a larger input limit is needed. Modality is not the issue because the contracts are plain text, which the model already accepts.
Question 7
An AI practitioner is comparing ways to customize a foundation model (FM) and must tell a compliance reviewer which approaches leave the model's weights unchanged. Which approaches meet that description? (Select TWO.)
More than one answer is right.
- In-context learning
- Retrieval Augmented Generation (RAG)
- Model distillation
- Continuous pre-training
- Fine-tuning
Show the answer
Answers: A. In-context learning; B. Retrieval Augmented Generation (RAG)
In-context learning and RAG only change what is placed in the prompt, so the weights stay as they are. Model distillation is the nearest distractor, but it fine-tunes a student model, which changes that model's weights.
Question 8
A developer at an insurance company uses a foundation model on Amazon Bedrock to decide whether claims meet a policy with five interdependent eligibility rules. The model often gives the wrong final decision. The developer rewrites the prompt to tell the model to work through each rule step by step before giving the decision. Which trade-off should the developer expect from this change?
- Longer outputs, which increase cost and latency
- Shorter outputs, which reduce cost but lower accuracy
- A need for labeled training data stored in Amazon S3
- Updated model weights, which require a new model deployment
Show the answer
Answer: A. Longer outputs, which increase cost and latency
Chain-of-thought prompting makes the model write out its reasoning, so outputs are longer and cost and latency rise. It changes only the prompt, so no weights are updated and no new deployment is needed.
Question 9
A product team launches a voice-controlled banking assistant. After launch, the team finds that people with speech impairments cannot complete basic tasks. A review shows that no users with speech impairments took part in the design research, and no recordings of their speech were in the training data. Which feature of responsible AI does this most directly violate?
- Robustness
- Safety
- Inclusivity
- Veracity
Show the answer
Answer: C. Inclusivity
Inclusivity means the system is designed for, and its data represents, the full range of people who will use it, and this group was left out of both. Robustness is the nearest distractor, but it concerns noisy, unexpected or adversarial input, not a user group that was never represented.
Question 10
A financial services company builds a chatbot on Amazon Bedrock. During a risk review, an auditor asks which security task belongs to the company under the AWS shared responsibility model. Which task is the company's responsibility?
- Preventing model providers from accessing Amazon Bedrock logs
- Protecting the global infrastructure that runs Amazon Bedrock
- Securing the physical facilities that host the foundation models
- Defining which IAM principals may invoke models and which guardrails apply
Show the answer
Answer: D. Defining which IAM principals may invoke models and which guardrails apply
Who may invoke models and which guardrails apply are configuration choices, and configuration is always the customer's side of the model. Protecting the global infrastructure that runs the service is the task AWS performs.
The full timed mock
65 questions in 90 minutes, with a pass mark of 70%. Each paper is drawn from a bank of questions written for the mock alone, and these 10 are left out of it. Full access opens it, with every level of the course.
LevelUpWith is independent of the organisation that sets this exam. These questions were written for this course; they are not taken from the exam.