Free Exam AI-901: Microsoft Azure AI Fundamentals Practice Questions

Sample Practice Questions

Try these 10 sample questions from the Exam AI-901: Microsoft Azure AI Fundamentals practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1What should a team examine in its training-data preparation to reduce the risk of discriminatory model outputs?

   1. A. Whether every training record has an identical text length
   2. B. Whether selection criteria and data reflect demographic bias
   3. C. Whether training records are sorted by their creation date
   4. D. Whether the model receives the entire dataset in one request

   Answer: B

   Explanation: Microsoft warns, "There's substantial risk that the data selection criteria, or the data itself reflects unconscious bias that may cause a model to produce discriminatory outputs." Inspecting selection criteria and data directly addresses that risk; record length, ordering, and request batching do not establish fairness.

   Reference: Responsible AI — Principles of responsible AI / Fairness

2. 2A Foundry chat assistant needs initial instructions defining its maintenance-support role and response style. Which message role is designed to carry these behavioral instructions?

   1. A. The user message role
   2. B. The system message role
   3. C. The assistant message role
   4. D. The tool message role

   Answer: B

   Explanation: Microsoft states, "you should use a system prompt to provide the model with instructions that guide its responses." The system role carries initial assistant behavior. User messages carry current requests, assistant messages represent replies, and tool messages carry tool results.

   Reference: Develop your first AI agent in Microsoft Foundry — Specify instructions in a system prompt

3. 3A hiring model evaluates applicants with similar job qualifications. Which outcome best reflects Microsoft's fairness principle?

   1. A. Recommendations differ solely because applicants belong to different age groups
   2. B. Recommendations favor the demographic group most common in the training set
   3. C. Recommendations remain comparable across otherwise similar applicant groups
   4. D. Recommendations follow the outcome historically assigned to each ethnic group

   Answer: C

   Explanation: Microsoft states, "AI systems should treat everyone fairly and avoid affecting similar groups differently." For similarly qualified applicants, comparable recommendations support this principle; demographic preferences and replication of historical group disparities do not.

   Reference: What is responsible AI? — Azure Machine Learning — Fairness and inclusiveness

4. 4In a role-based chat request, which role should contain the end user's current question?

   1. A. system
   2. B. assistant
   3. C. user
   4. D. tool

   Answer: C

   Explanation: The guide places the text "The user's message goes here." in the user-role message. The current question belongs in user; system defines behavior, assistant represents replies, and tool represents tool results.

   Reference: Work with chat completion models — Work with the Chat Completions API / basic chat format

5. 5An AI model has a low overall error rate. What additional assessment can reveal reliability problems hidden by that average?

   1. A. Compare error rates for relevant subsets of the evaluation data
   2. B. Report the same aggregate error rate with more decimal places
   3. C. Count how many predictions the model can return per second
   4. D. Replace the evaluation inputs with copies of training inputs

   Answer: A

   Explanation: Microsoft describes identifying "cohorts (subsets) of data with a higher error rate than the overall benchmark." Subset analysis can reveal concentrated failures that the overall average hides. Precision formatting, throughput, and reusing training examples do not provide that comparison.

   Reference: What is responsible AI? — Azure Machine Learning — Reliability and safety / error analysis

6. 6A developer wants a chat model to imitate a particular answer format using few-shot prompting. Which prompt design applies this technique?

   1. A. Include example user questions and corresponding assistant answers before the new question
   2. B. Describe the desired answer style in instructions without providing demonstrations
   3. C. List example user questions but provide none of the corresponding example answers
   4. D. Provide sample assistant answers without the example inputs they are meant to answer

   Answer: A

   Explanation: Microsoft says, "You can now include a series of messages between the user and the assistant in the prompt as few-shot examples." Paired demonstrations show the desired mapping and format. Instructions alone, unanswered questions, and unpaired answers do not provide those input-output examples.

   Reference: Work with chat completion models — Use few-shot learning / Python

7. 7Why should an application mitigate the risk of incorrect AI outputs even after its model performs well in testing?

   1. A. High average test accuracy proves that rare harmful cases are already covered
   2. B. Passing the test set establishes that production inputs cannot cause errors
   3. C. A successful model test makes later output-risk controls unnecessary
   4. D. AI uses probabilistic models that can still produce errors

   Answer: D

   Explanation: Microsoft states, "AI is based on probabilistic models, it is not infallible." Average accuracy and passing a test set do not prove coverage of every production condition or eliminate the need to mitigate errors.

   Reference: Responsible AI — Principles of responsible AI / Reliability and safety

8. 8A chat application must answer from a large collection of company documents without placing the entire collection into every prompt. Which approach does Microsoft's grounding guidance describe?

   1. A. Supply document titles as context while omitting the supporting document passages
   2. B. Supply the same fixed excerpt for every query without selecting for relevance
   3. C. Retrieve relevant passages but discard them before constructing the model input
   4. D. Retrieve relevant information at query time and include it as context

   Answer: D

   Explanation: Microsoft says to "retrieve the most relevant information at query time." Including that retrieved content supplies query-specific evidence. Titles alone, an unselected fixed excerpt, and passages discarded before the request do not supply the required relevant document evidence.

   Reference: Work with chat completion models — Use data for grounding

9. 9Which behavior most directly demonstrates reliability and safety when an AI system encounters an unexpected input condition?

   1. A. Publishing the model's feature list alongside the result
   2. B. Handling the condition safely instead of continuing a hazardous action
   3. C. Retaining every input indefinitely for possible future model training
   4. D. Assuming the condition matches the most common training example

   Answer: B

   Explanation: Microsoft says systems should "function as designed, respond safely to unexpected conditions, and resist harmful manipulation." Safe handling addresses the unexpected condition; disclosure alone does not control it, and unlimited retention or unverified assumptions do not make the response safe.

   Reference: What is responsible AI? — Azure Machine Learning — Reliability and safety

10. 10During model-playground testing, a developer wants to test a prompt without earlier conversation turns influencing it. Which action clears that conversation history?

    1. A. Ask the model to disregard history in a new prompt
    2. B. Select New chat in the chat pane
    3. C. Save the current model configuration as an agent
    4. D. Change the assistant's style instruction for the same chat

    Answer: B

    Explanation: The official exercise says, "use the New chat button to restart the conversation. This removes all conversation history." A request to disregard history, saving configuration, or changing a behavioral instruction does not itself delete the existing conversation turns.

    Reference: Develop your first AI agent in Microsoft Foundry — Chat with the model