Microsoft AI-901 Practice Test 39

Question 380

A voice assistant reads confidential medical results aloud over a shared reception speaker. How should the application protect the user?

Require appropriate identity, consent and private audio-delivery context before speaking protected results.
Disable text display but continue unrestricted voice playback.
Use a softer voice while keeping the same public speaker.
Assume that because the patient requested results once, any nearby listener may hear them.

Explanation

Correct Answer

Confirm the patient’s identity and permission, check whether the selected speaker is private, and use a consent-aware audio-delivery path before reading protected results aloud.

Incorrect Answers

Hiding the text display does not stop people in the waiting area from hearing a shared speaker. Confidentiality includes audio channels.

A softer voice may still be audible and leaves the same unauthorized listeners present. Volume is not an authorization mechanism.

A prior request for results does not imply consent to broadcast medical details to everyone nearby. Verify the current delivery context.

Question 381

A newly offered synthesized voice sounds good in a demo but is not documented for the target locale. What should developers verify before release?

Assume every voice supports every language because it can output speech.
Check only that the synthesis endpoint returns HTTP 200 for an unrelated demo sentence, without testing target-locale support or actual pronunciation.
Change only the user interface locale without changing the configured synthesis voice.
Supported voice/locale pairing and pronunciation on representative target-language text.

Explanation

Correct Answer

Validate the target voice/locale combination against current Speech support and listen to representative actual-language prompts. Check pronunciation and any SSML features needed before deployment.

Incorrect Answers

Voices have documented language and locale support; the ability to generate speech in one language does not prove accurate pronunciation elsewhere.

API success on a different phrase does not establish support for the release language or correct reading of critical localized terms. Test the approved voice against representative text.

Changing only the displayed interface language does not switch the configured synthesis voice or its pronunciation engine.

Question 382

A hardware-support app already uses a supported image-capable model to inspect several photos of a laptop hinge. In one photo a dark line may be either a casing crack or a reflection. A warranty replacement is expensive. What should the app do before approving the replacement?

Compare the relevant hinge region across the original photos, request a clearer view if still ambiguous, and route uncertain damage findings for authorized review.
Accept a high crack-confidence score from the darkest photograph without comparing the source images for reflections.
Use the laptop model’s historical hinge failure rate as proof that this specific unit has a visible crack.
Apply generative image enhancement to create a more obvious crack and treat the synthesized pixels as the original customer evidence.

Explanation

Correct Answer

An ambiguous dark line is not yet proof of a physical fracture. Compare source-image regions across captures with different lighting, retain uncertainty, obtain clearer evidence when needed and require the appropriate reviewer before committing a high-cost warranty action.

Incorrect Answers

A high score on one ambiguous frame can reflect glare or compression rather than a crack. Comparing source pixels and additional captures checks whether the apparent line persists as physical damage rather than a transient optical effect.

A population failure rate can inform inspection priority but cannot establish that this particular hinge is cracked. Warranty approval depends on the submitted unit’s visual evidence and the governing business policy.

Generative enhancement can invent detail that was not visible in the source photographs. Synthetic pixels are not authenticated evidence of a physical break and cannot replace direct source-image validation.

Question 383

A programmer includes the URL `https://.../photo.png` only as plain text in a prompt and assumes a text-only model will inspect the photograph. What must be changed?

Use a descriptive variable name `photo_contents` while still sending only a URL string.
Supply image bytes or a supported image-content reference using a model and API that actually process visual input.
Replace the URL with a different file extension without fetching the image.
Increase the response limit so the model has more tokens to guess what is at the URL.

Explanation

Correct Answer

Supply actual image bytes or a provider-supported image URL/content reference to an image-capable model. Confirm the endpoint can access the source and that the image is processed rather than merely mentioned.

Incorrect Answers

A descriptive variable name does not change the model’s input schema or make it fetch a remote photograph. The request still contains only a string.

Renaming a file or URL does not reveal its visual content to a text-only model and may break media decoding entirely.

Extra generation tokens allow a longer answer, not inspection of unseen pixels. Any specific visual details could be hallucinated.

Question 384

An image-inspection assistant says a photo contains three identical valves. A labeled reviewer can see four. Which evaluation is necessary before trusting automated counts?

Trust the image-level label "valve" because class presence proves object quantity.
Compare predictions to labeled per-instance detections on representative images and inspect missed/duplicate objects.
Use the assistant's prose confidence as an exact count validator.
Regenerate an image with four valves and use that synthetic frame as the source.

Explanation

Correct Answer

Compare model detections to annotated individual valves across representative images, including occlusion and overlapping boxes. Separate misses from duplicated detections to diagnose why the reported count is three rather than four.

Incorrect Answers

An image-level classification label says valves are present but supplies no instance count. The same label could be returned for one or ten valves.

A model’s prose confidence is not calibrated evidence of an exact object count. Reconcile predicted instances against ground-truth annotations.

A generated four-valve picture is not the physical inspection source. It cannot establish whether the original photo was counted correctly.

Question 385

A screenshot of a financial chart contains an axis labeled "millions," but a multimodal model reports values as individual dollars. Which check should block the response?

Verify axis labels, scale and units against the original screenshot before presenting numerical conclusions.
Accept the model's units because the chart trend direction seems right.
Infer units from the surrounding report's prose without checking the image label.
Increase image generation quality to refine the chart's appearance.

Explanation

Correct Answer

Inspect the chart’s explicit “millions” label, numerical ticks and any scaling before reporting amounts. A plotted value of 2 may represent two million dollars, not two individual dollars.

Incorrect Answers

Correct trend direction does not guarantee the numbers or units are right. An axis-scale error can change financial magnitude by a factor of one million.

Surrounding prose may be incomplete or describe a different figure; the uploaded screenshot itself contains the controlling axis label.

Image-generation quality changes aesthetics, not the accuracy of interpreting an existing chart. It cannot repair numerical unit analysis.

Question 386

A repair technician submits a blurry image of a serial plate, and the vision model invents two missing digits. What should be done instead?

Mark the guessed serial as final because the visual image is at least partially clear.
Use the language model's self-confidence to select a guessed suffix.
Append the two most common digits for that model family.
Request a sharper close-up or higher-resolution crop and label unreadable characters uncertain until verified.

Explanation

Correct Answer

Ask for a clearer source crop or recapture and explicitly mark unresolved characters. Cross-check against a trusted asset registry before using the serial for warranty or safety records.

Incorrect Answers

A partly legible serial cannot be treated as a complete verified identifier when two characters were invented. This risks misregistering the repaired asset.

Model confidence describes its own uncertain inference, not observed strokes hidden by blur. It does not validate the guessed suffix.

Common suffix statistics describe other units and cannot uniquely identify this specific plate. Guessing may match the format while pointing to another device.

Question 387

A mobile application sends enormous raw photographs to a supported vision endpoint and receives size-limit errors. What preprocessing is appropriate?

Retry the same oversized payload indefinitely.
Discard the photo's corner containing the serial label just to make the request smaller.
Resize or crop within documented limits while preserving the task-critical region and source traceability.
Re-encode the photograph to JPEG and assume it now meets the API contract without checking image dimensions, remaining payload bytes or label legibility.

Explanation

Correct Answer

Resize or crop to supported pixel and file-size limits while retaining the region of interest and source metadata. Confirm any label remains legible after preprocessing.

Incorrect Answers

Retrying identical oversize image bytes repeats a deterministic upload constraint and does not change the endpoint’s maximum payload size.

Blind cropping may discard the serial label, the only region needed for verification. Compression must preserve task-relevant evidence.

Encoding may reduce bytes, but the endpoint can impose dimension and payload limits, while aggressive compression damages fine label text. Verify the actual result against supported size limits and the task-critical region.

Question 388

An insurance claims photo includes a customer's driver's license beside a damaged product. The image classifier only needs the product condition. What privacy gate is needed?

Crop or redact irrelevant personal identifiers and enforce authorized image access before inference and logging.
Run OCR redaction only after sending the full license image through every model in the pipeline.
Enable full-image logging for every claims model while restricting only who can view the final customer answer.
Crop the finished report around the damaged product but retain the unredacted license photograph in unrestricted request logs.

Explanation

Correct Answer

The claims classifier requires product damage, not a driver’s license. Remove or mask irrelevant identity pixels before model input and restrict access to any retained original and traces.

Incorrect Answers

Late redaction may protect some display text but the license pixels have already entered unnecessary model contexts or traces. Minimize protected source content before inference where feasible.

Logging every raw driver’s license to broad model traces creates exposure beyond the damaged-product task. Privacy controls apply before storage and across internal systems, not just in the UI.

A cropped report does not remove the raw personal document from logs. The source and diagnostic data need independent authorization, masking and retention controls.

Question 389

An inspector asks a multimodal app, 'Is there a crack in this wall?' but the client sends an empty image field and only a textual description of the inspection task. How should the app answer?

Generate a plausible photograph of a cracked wall and use that synthetic picture as the inspection record.
Detect the missing image and request the actual wall photograph or another verified inspection record before assessing the crack.
List common wall-crack patterns and claim they are visible in the site photograph, even though no image arrived.
Report that the wall has no visible cracks because the image-upload field contains no pixels.

Explanation

Correct Answer

Without the actual wall image or a verified inspection record, no visual claim about this building is supported. The client must identify the missing input and request evidence before deciding whether a crack is present.

Incorrect Answers

A generated wall photo is not evidence from the actual building. Using synthesized damage as if observed could create false maintenance decisions and mislead the inspector.

Typical crack patterns are general knowledge, not observations about this wall. Reporting them as visible would falsely imply access to an inspection image.

An empty payload means no visual inspection occurred; it does not establish that the wall is free from cracks. The result must be missing evidence, not a negative finding.

Question 390

A visual alarm-monitoring assistant flags a red indicator as an active alarm, but sunlight can make an inactive orange lamp look red. Select TWO appropriate validation steps.

Declare every red-looking pixel an active alarm without checking which lamp it belongs to.
Test representative lighting conditions and compare predictions to labeled actual alarm states.
Correlate the visual signal with the authoritative device status before triggering a consequential response.
Generate an idealized red lamp image and treat it as the camera input.
Increase response verbosity so the model explains its color guess more persuasively.

Explanation

Correct Answers

Evaluate images across daylight, glare and lamp states against labeled actual alarm outcomes. This measures false activations caused by orange lamps appearing red.

Before a consequential alert, reconcile the camera classification with the authoritative device telemetry or alarm-state service so appearance does not substitute for actual status.

Incorrect Answers

Color alone is ambiguous under changing illumination and does not verify that the observed pixel belongs to the correct alarm indicator.

A newly generated ideal lamp omits the real lighting and equipment context. It is unsuitable as ground truth for this physical alarm.

Additional explanation words cannot turn a mistaken color perception into a reliable alarm signal. Outcome validation is required.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests

Microsoft AI-901 Practice Test 37 - 10 Questions (Last Reviewed 08-Oct-2026)
Microsoft AI-901 Practice Test 38 - 10 Questions (Last Reviewed 07-Oct-2026)
Microsoft AI-901 Practice Test 39 - 10 Questions (Last Reviewed 06-Oct-2026)
Microsoft AI-901 Practice Test 40 - 10 Questions (Last Reviewed 05-Oct-2026)
Microsoft AI-901 Practice Test 41 - 10 Questions (Last Reviewed 04-Oct-2026)

Pass Your Microsoft AI-901 Exam Easy!

Question 380

Explanation

Correct Answer

Incorrect Answers

Question 381

Explanation

Correct Answer

Incorrect Answers

Question 382

Explanation

Correct Answer

Incorrect Answers

Question 383

Explanation

Correct Answer

Incorrect Answers

Question 384

Explanation

Correct Answer

Incorrect Answers

Question 385

Explanation

Correct Answer

Incorrect Answers

Question 386

Explanation

Correct Answer

Incorrect Answers

Question 387

Explanation

Correct Answer

Incorrect Answers

Question 388

Explanation

Correct Answer

Incorrect Answers

Question 389

Explanation

Correct Answer

Incorrect Answers

Question 390

Explanation

Correct Answers

Incorrect Answers

Microsoft Azure AI Fundamentals (AI-901) Practice Tests