McAfee Secure

Exam Code: AI-901

Exam Name: Microsoft Azure AI Fundamentals

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Fundamentals

Microsoft AI-901 Questions & Answers

Study with Up-To-Date REAL Exam Questions and Answers from the ACTUAL Test

166 Questions & Answers with Testing Engine

"Microsoft Azure AI Fundamentals Exam", also known as AI-901 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-901 Exam Engine. Your AI-901 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-901 Practice Test 4

Question 30

Troubleshooting report: The explanation equates one token to one full English word in all cases. Which correction is MOST directly targeted to this failure?

  1. Chunk and retrieve relevant source sections before inference
  2. Demonstrate tokenization differences across words and scripts
  3. Implement and test fixed numeric rules in application code
  4. Re-embed the corpus or use a versioned compatible index

Explanation

Correct Answer

B:

This correction directly addresses the reported failure. Tokenization splits text into subword or other units for model processing. The targeted change is: Demonstrate tokenization differences across words and scripts.

Incorrect Answers

A:

A finite context window limits how much material can be processed at once. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Chunk and retrieve relevant source sections before inference” does not target that failure.

C:

Generative models are not a substitute for auditable deterministic computation. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Implement and test fixed numeric rules in application code” does not target that failure.

D:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Re-embed the corpus or use a versioned compatible index” does not target that failure.

Question 31

Troubleshooting report: Training examples treat unfamiliar slang as automatically malicious. Which correction is MOST directly targeted to this failure?

  1. Add subgroup and matched-pair fairness tests before approval
  2. Require a data query or report unavailability
  3. Archive the actual decision and versioned provenance with privacy minimization
  4. Rebalance safety tests and adjust erroneous dialect associations

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Fairness requires discriminating actual harm from linguistic variety. The targeted change is: Rebalance safety tests and adjust erroneous dialect associations.

Incorrect Answers

A:

Fairness checks compare materially equivalent cases rather than raw group averages alone. In the stated case, troubleshooting report: Training examples treat unfamiliar slang as automatically malicious; “Add subgroup and matched-pair fairness tests before approval” does not target that failure.

B:

A numeric-looking answer is not proof that a measure was computed. In the stated case, troubleshooting report: Training examples treat unfamiliar slang as automatically malicious; “Require a data query or report unavailability” does not target that failure.

C:

Auditability requires the original decision evidence and version identifiers; regenerating model text is not guaranteed to be identical. In the stated case, troubleshooting report: Training examples treat unfamiliar slang as automatically malicious; “Archive the actual decision and versioned provenance with privacy minimization” does not target that failure.

Question 32

An engineering review documents this requirement: A support agent receives conflicting instructions from a user and the app policy. Select the most appropriate solution.

  1. Use the supported run or response status and verify its terminal state
  2. Call the deployed model through an authenticated application backend
  3. Preserve the higher-priority application constraints in the message hierarchy
  4. Handle tool timeout as a typed error and preserve partial workflow state

Explanation

Correct Answer

C:

Developer and system constraints govern allowed operations over user requests. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Submission is not the same as successful completion. That is not sufficient here: Developer and system constraints govern allowed operations over user requests.

B:

Backend controls protect credentials, quotas and data policy. That is not sufficient here: Developer and system constraints govern allowed operations over user requests.

D:

Timeout handling should not be reported as a successful action. That is not sufficient here: Developer and system constraints govern allowed operations over user requests.

Question 33

An Azure AI solution uses the following approved control: Minimize unnecessary identifying visual data and enforce approved access. What observable result should pass QA?

  1. A restricted summary omits or masks nonessential identifying features
  2. The user sees pending or failed status when no payment receipt exists
  3. Eligible memory entries disappear from active stores after the verified request
  4. An incident simulation reaches a named owner with clear containment authority

Explanation

Correct Answer

A:

Privacy applies to image content as well as text transcripts. The expected acceptance evidence is: A restricted summary omits or masks nonessential identifying features. This observation directly tests the stated behavior.

Incorrect Answers

B:

Generated text is not evidence of external side effects. This observation does not establish the required result: A restricted summary omits or masks nonessential identifying features.

C:

Accountability requires evidence of actual data lifecycle actions. This observation does not establish the required result: A restricted summary omits or masks nonessential identifying features.

D:

An AI model cannot be its own accountable incident responder. This observation does not establish the required result: A restricted summary omits or masks nonessential identifying features.

Question 34

Troubleshooting report: The team updates query embeddings but leaves stored embeddings from the old model. Which correction is MOST directly targeted to this failure?

  1. Re-embed the corpus or use a versioned compatible index
  2. Replace it with a suitable text-analysis model or service
  3. Implement and test fixed numeric rules in application code
  4. Use a speech synthesis endpoint and compatible output format

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. Incompatible vector spaces can make nearest-neighbor rankings meaningless. The targeted change is: Re-embed the corpus or use a versioned compatible index.

Incorrect Answers

B:

Different model tasks require compatible outputs and labeled evaluation data. In the stated case, troubleshooting report: The team updates query embeddings but leaves stored embeddings from the old model; “Replace it with a suitable text-analysis model or service” does not target that failure.

C:

Generative models are not a substitute for auditable deterministic computation. In the stated case, troubleshooting report: The team updates query embeddings but leaves stored embeddings from the old model; “Implement and test fixed numeric rules in application code” does not target that failure.

D:

Speech synthesis generates audio from text rather than transcribing audio. In the stated case, troubleshooting report: The team updates query embeddings but leaves stored embeddings from the old model; “Use a speech synthesis endpoint and compatible output format” does not target that failure.

Question 35

An Azure-based AI team must satisfy this objective: An AI video summarizer may expose faces of bystanders who never consented. Which design choice is best?

  1. Enforce the monetary threshold in the backend and require human authorization above it
  2. Represent uncertainty explicitly and refer to approved teaching sources
  3. Use validated data retrieval and source-backed numerical responses
  4. Minimize unnecessary identifying visual data and enforce approved access

Explanation

Correct Answer

D:

Privacy applies to image content as well as text transcripts. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

A prompt-level request cannot override a transaction policy. That is not sufficient here: Privacy applies to image content as well as text transcripts.

B:

Honesty requires acknowledging when a claim cannot be substantiated. That is not sufficient here: Privacy applies to image content as well as text transcripts.

C:

A numeric-looking answer is not proof that a measure was computed. That is not sufficient here: Privacy applies to image content as well as text transcripts.

Question 36

Troubleshooting report: The same examples are used for training and claimed as unbiased testing. Which correction is MOST directly targeted to this failure?

  1. Split data by appropriate entities and evaluate on unseen cases
  2. Evaluate mean absolute error or another appropriate regression metric
  3. Use early stopping and investigate excessive model capacity
  4. Keep external content in lower-trust user/tool data fields

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. A model can memorize training examples without generalizing. The targeted change is: Split data by appropriate entities and evaluate on unseen cases.

Incorrect Answers

B:

Regression predicts continuous quantities under a defined error metric. In the stated case, troubleshooting report: The same examples are used for training and claimed as unbiased testing; “Evaluate mean absolute error or another appropriate regression metric” does not target that failure.

C:

More optimization on training data can worsen generalization. In the stated case, troubleshooting report: The same examples are used for training and claimed as unbiased testing; “Use early stopping and investigate excessive model capacity” does not target that failure.

D:

Instruction roles help distinguish application policy from user task text. In the stated case, troubleshooting report: The same examples are used for training and claimed as unbiased testing; “Keep external content in lower-trust user/tool data fields” does not target that failure.

Question 37

The approved design action is: Stop or regularize training based on held-out validation behavior. Which test result would BEST establish that it worked?

  1. The returned text includes the photographed serial number
  2. The holdout report identifies missed positives and false alarms explicitly
  3. Early stopping selects a model before validation performance degrades
  4. A regression test records model version and key generation settings

Explanation

Correct Answer

C:

More optimization on training data can worsen generalization. The expected acceptance evidence is: Early stopping selects a model before validation performance degrades. This observation directly tests the stated behavior.

Incorrect Answers

A:

OCR produces readable text from pixels. This observation does not establish the required result: Early stopping selects a model before validation performance degrades.

B:

High accuracy can conceal failure to detect rare positive events. This observation does not establish the required result: Early stopping selects a model before validation performance degrades.

D:

Reproducibility depends on more than the visible prompt text. This observation does not establish the required result: Early stopping selects a model before validation performance degrades.

Question 38

A project needs the following AI behavior: A text-to-image output contains unwanted unsafe material. What should the designer implement?

  1. Use a supported dense-captioning capability where available
  2. Use image classification for the requested whole-image class
  3. Select appropriate separate visual features for OCR, tagging and people detection
  4. Apply supported image safety filtering and review before distribution

Explanation

Correct Answer

D:

Generative media needs policy controls in addition to prompt instructions. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Dense captions describe localized regions beyond a single global caption. That is not sufficient here: Generative media needs policy controls in addition to prompt instructions.

B:

Image-level classification is distinct from object localization. That is not sufficient here: Generative media needs policy controls in addition to prompt instructions.

C:

Different visual outputs describe text, scene classes and detected people. That is not sufficient here: Generative media needs policy controls in addition to prompt instructions.

Question 39

The approved design action is: Send the image in a supported multimodal message along with the question. Which test result would BEST establish that it worked?

  1. Each cited claim resolves to a real approved source chunk
  2. The difference can be tied to an explicit configuration mismatch
  3. The service processes role-tagged messages with the expected application behavior
  4. The answer mentions relevant visual details unique to the uploaded image

Explanation

Correct Answer

D:

The model must receive the actual visual content to reason about it. The expected acceptance evidence is: The answer mentions relevant visual details unique to the uploaded image. This observation directly tests the stated behavior.

Incorrect Answers

A:

Explicit response structure makes evidence easier to verify. This observation does not establish the required result: The answer mentions relevant visual details unique to the uploaded image.

B:

Environment drift can change behavior without code edits. This observation does not establish the required result: The answer mentions relevant visual details unique to the uploaded image.

C:

Chat APIs use structured messages rather than concatenating untyped command strings. This observation does not establish the required result: The answer mentions relevant visual details unique to the uploaded image.

Question 40

An Azure AI solution uses the following approved control: Run the required text analysis operations separately and preserve distinct result types. What observable result should pass QA?

  1. The generated audio plays through the selected output device
  2. The response contains both an attitude label and typed entity spans
  3. The output audio reads the approved sentence intelligibly
  4. Mic capture ceases and the app no longer receives audio frames

Explanation

Correct Answer

B:

Sentiment polarity and named entity spans answer different questions. The expected acceptance evidence is: The response contains both an attitude label and typed entity spans. This observation directly tests the stated behavior.

Incorrect Answers

A:

Successful synthesis does not guarantee local audio is routed to a speaker. This observation does not establish the required result: The response contains both an attitude label and typed entity spans.

C:

Speech synthesis generates audio waveforms from supplied text. This observation does not establish the required result: The response contains both an attitude label and typed entity spans.

D:

User control of capture must be enforced at the device or application layer. This observation does not establish the required result: The response contains both an attitude label and typed entity spans.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests