Microsoft AI-901 Practice Test 40

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.

Question 391

An expense-review agent reads a ticket photograph and returns an invoice date. A reviewer sees that the number came from the adjacent price field. What extraction check is missing?

Trust a fluent model explanation over the photographed layout.
Validate the date against OCR source spans and nearby labels before using it as a calendar value.
Use the image's creation timestamp as the invoice date by default.
Choose any digit sequence that meets a date regex without checking the label.

Explanation

Correct Answer

Inspect OCR spans and neighboring labels, validate date formatting and compare with the actual invoice image before populating a calendar field. This catches a price misread as a date.

Incorrect Answers

Fluent prose cannot change which printed field supplied the value. The visual layout and labels, not response confidence, determine whether a number represents a date.

Image creation timestamp records capture or file processing, not necessarily the invoice’s legally relevant date.

A digit string can match a date regex while coming from an amount, account or reference number. Semantic field mapping needs source evidence.

Question 392

A warehouse photo shows several identical boxes, and a worker asks "Is the box nearest the exit damaged?" The model answers about a different box. What is the needed control?

Select the first detected parcel box with a high confidence score without locating the exit in the original photograph.
Use a whole-image damage score and infer that it describes the box closest to the exit.
Ground the referent in spatial cues or a selected object region and verify the targeted box before classifying damage.
Choose the highest-confidence parcel detection and report its condition without resolving which region is nearest the exit.

Explanation

Correct Answer

Identify the exit in the actual image, localize the referenced box and verify its region before assessing damage. Object-level grounding prevents answers about a neighboring box.

Incorrect Answers

A high-confidence parcel detection may refer to a different box. The request identifies the parcel by its proximity to the exit, so that spatial relationship must be checked.

A global image classification does not bind damage to a specific parcel. It can report a damaged box elsewhere while the requested box is undamaged.

Detection confidence means the system recognized a parcel, not that the parcel satisfies the user’s spatial reference. Explicit object grounding is required.

Question 393

A production-line camera must draw boxes around individual defects. A team proposes an image-captioning model that only returns prose. Which capability should be compared?

Use whole-frame defect classification and draw one bounding box around the entire camera image.
Object detection or localization that returns spatial coordinates for the actual defects.
Configure a segmentation pipeline only for one whole-image defect/no-defect output rather than instance coordinates.
Run OCR to search for printed fault labels on the equipment and treat those words as coordinates of visible defects.

Explanation

Correct Answer

Use an object detection or localization workflow that produces bounding boxes or masks for each actual defect. Evaluate spatial precision against labeled images rather than free-form text.

Incorrect Answers

An image-level defect probability cannot localize which separate components contain cracks. A box around the whole frame is not the requested per-defect geometry.

A segmentation-capable family can localize defects, but this proposed configuration discards instance boundaries and does not return the boxes needed for separate actual defects.

OCR can read printed labels but does not localize a physical fracture unless a relevant spatial detection workflow identifies the crack itself. Text locations are not defect boxes.

Question 394

A multimodal inspection model is uncertain whether a safety guard is installed. Maintenance wants to automatically certify the machine as safe. What should the system do?

Approve the machine when the response sounds confident despite poor image evidence.
Generate a reference image of an installed guard and assume the inspected machine matches.
Require additional visual evidence or qualified human review rather than certify an uncertain safety-critical state.
Mark every unrecognizable component as installed to avoid delays.

Explanation

Correct Answer

Ask for clearer photographs, independent instrument evidence or qualified inspection when visual support is uncertain. Do not certify a safety-critical state on an unverified model guess.

Incorrect Answers

A fluent answer about a guard can still be wrong when key pixels are occluded or blurred. Automatic safety certification requires verified evidence.

A generated reference picture illustrates a safe guard but does not prove it exists on the actual machine.

Automatically treating unknown components as installed is a fail-open rule that can permit hazardous machinery into service.

Question 395

An image-enabled Foundry deployment is replaced with a model revision that accepts different image formats. What must the client verify before switching users?

Delete prior image examples so there is no baseline.
Assume any version that returns text accepts every previous image payload.
Supported input schema, file formats and representative visual-result regressions on the target version.
Switch image formats by renaming extensions only, without verifying byte encoding.

Explanation

Correct Answer

Check supported media formats, image-content schema, endpoints and sample visual tasks against the new deployment. Confirm actual pixel analysis and source-result parity before switching users.

Incorrect Answers

Deleting earlier test images removes the evidence needed to detect regression in formats, performance or safety on the replacement model.

A text response does not prove that the model interpreted the actual image payload; the new version might reject or ignore the earlier request shape.

Renaming files cannot convert unsupported media codecs or change the bytes required by the image processor.

Question 396

A user uploads a chart whose vertical axis is explicitly logarithmic, with labeled ticks at 1, 10, 100 and 1,000. A vision answer says the halfway point between the 10 and 100 ticks represents 55. Which interpretation check is essential?

Read the pixel spacing accurately but interpolate 10 and 100 arithmetically, because the labels are equally spaced visually.
Compute the midpoint of the log10 tick values as 1.5, then report the plotted quantity as 1.5 instead of converting back to the original units.
Read the axis-scale label and use the corresponding logarithmic interpretation; a geometric midpoint between 10 and 100 represents roughly 31.6, not the arithmetic midpoint 55.
Recreate the plotted series on a linear axis with the original labels, then read 55 from the reconstructed image rather than the supplied chart.

Explanation

Correct Answer

Read the log-scale tick marks and compute a geometric midpoint: sqrt(10 × 100) is approximately 31.6. Verify that the reported value matches the image axis rather than averaging endpoints.

Incorrect Answers

Equal vertical distances on a logarithmic axis correspond to equal ratios, not equal numeric increments. Arithmetic midpoint 55 is not the plotted halfway value.

The halfway log-coordinate is 1.5, but the vertical data value is 10 to that power, about 31.6. The inverse transformation is required before reporting the quantity.

Changing the axis transform changes the meaning of the plotted position; an inferred linear replot cannot supersede the original log-scale chart when estimating its values.

Question 397

A public transit app generates alternate text for photographs used by blind travelers. How should the team evaluate its descriptions?

Measure caption readability alone on a set of generated stock photographs, without checking the original transit scenes.
Test factual coverage and unsupported detail rates with real source images and accessibility users.
Measure object-label precision on a small internal sample but do not test whether the alternative text communicates the wayfinding detail travelers need.
Require every photo to have one generic accessibility label, and treat successful screen-reader playback as evidence that its content was described correctly.

Explanation

Correct Answer

Representative source images reveal missed objects and invented guidance, while real assistive-technology users can judge whether the description supports the intended travel task.

Incorrect Answers

Readability is useful but does not establish that the caption identifies the correct route sign, obstacle or vehicle on the actual photograph. Evaluation needs original transit evidence and travelers’ access needs.

Object-label precision is only one component. A caption can name objects correctly yet omit the relationship or direction that a blind traveler needs to navigate safely.

Screen-reader compatibility means text can be announced, not that the announced words faithfully describe the particular image. A generic label omits meaningful visual information.

Question 398

A user uploads a PDF containing a wiring diagram but the selected endpoint accepts only image parts. How should the app handle the file?

Pass only the document title and assume the model can inspect the diagram.
Base64-encode the complete PDF and paste it into a text-only message, rather than using a documented document or rendered-page image input.
Rename the PDF extension to `.jpg` without converting the content.
Use an approved PDF/document-capable path or render relevant pages as images within the supported contract.

Explanation

Correct Answer

Use a supported PDF/document model or render selected pages into accepted images, preserving page numbers and image quality. The visual client can then inspect the actual diagram.

Incorrect Answers

A document filename does not expose the diagram geometry, symbols or wire connections. The model needs supported visual input.

Encoding the PDF as ordinary text does not guarantee that an image-only endpoint will parse the document or expose the diagram pixels to visual reasoning.

Renaming PDF bytes to .jpg does not render its pages or produce valid image data for a vision API.

Question 399

A compliance reviewer uploads a collage of twelve facility photographs. The model confuses a missing guard in facility A with machinery in facility B. Which input strategy improves grounding?

Apply one image-wide "safe" class to the entire collage.
Regenerate a representative facility photo and evaluate that instead.
Combine all photos with no location labels to maximize context.
Separate or explicitly region-label the images and correlate each output to the correct facility/source.

Explanation

Correct Answer

Process photos separately or annotate each region with verified facility and source IDs, then attach every finding to the corresponding photo. This prevents misassigning a guard defect.

Incorrect Answers

A single collage-level “safe” classification loses which facility and piece of machinery has the defect. It cannot support facility-specific action.

A generated average facility image is not a record of any of the twelve inspections. Using it hides the observed missing guard.

Unlabeled merged inputs promote cross-photo confusion because visually similar machines can belong to different sites.

Question 400

A marketing team already has an approved text-to-image Foundry deployment. Its app submits a valid campaign brief, then displays “Image ready” whenever the API returns HTTP success—even when the response contains no image bytes. What must be checked before delivering the artwork?

Consider a successful request log entry sufficient to deliver the artwork, even if the response body has no image payload.
Generate a unique media filename before calling the tool and treat that filename as proof that the image was created.
Inspect the supported image-generation result or tool-call output for actual image data and success state, then persist and validate the resulting media before showing it as delivered.
Use the natural-language summary from the model as the generated image when image generation did not return media bytes.

Explanation

Correct Answer

Image-generation APIs and tools return specific output items or encoded media on success. The client should validate the expected result type, decode/store image bytes, link them to the request and report failures honestly.

Incorrect Answers

HTTP or orchestration success can mean that a request was accepted or a response envelope was created. It does not establish that the expected image payload was produced or stored.

An output filename is application metadata. Assigning it before the service generates media does not create image bytes or prove that the resulting asset matches the current request.

A textual description may explain an image request but is not the requested visual asset. Treat a missing image result as incomplete or failed rather than claiming an illustration was delivered.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests

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)
Microsoft AI-901 Practice Test 42 - 10 Questions (Last Reviewed 03-Oct-2026)

Pass Your Microsoft AI-901 Exam Easy!

Question 390

Explanation

Correct Answers

Incorrect Answers

Question 391

Explanation

Correct Answer

Incorrect Answers

Question 392

Explanation

Correct Answer

Incorrect Answers

Question 393

Explanation

Correct Answer

Incorrect Answers

Question 394

Explanation

Correct Answer

Incorrect Answers

Question 395

Explanation

Correct Answer

Incorrect Answers

Question 396

Explanation

Correct Answer

Incorrect Answers

Question 397

Explanation

Correct Answer

Incorrect Answers

Question 398

Explanation

Correct Answer

Incorrect Answers

Question 399

Explanation

Correct Answer

Incorrect Answers

Question 400

Explanation

Correct Answer

Incorrect Answers

Microsoft Azure AI Fundamentals (AI-901) Practice Tests