Microsoft AI-901 Practice Test 42

Question 410

A generated safety diagram depicts a clearance distance of "12 cm," but engineers suspect the geometry corresponds to a different scale. What should happen?

Ignore the mismatch and trust the surrounding caption.
Verify dimensions against approved technical specifications and measured geometry before publishing the diagram.
Regenerate until the label looks more convincing without consulting specifications.
Run OCR to confirm that the printed label reads “12 cm,” then assume the generated diagram’s geometry matches that physical clearance.

Explanation

Correct Answer

Compare labeled distance against approved engineering specifications and measured geometry before publication. A plausible 12 cm label on an inaccurate diagram can mislead maintenance.

Incorrect Answers

A convincing caption does not establish scale correctness or the clearance required for safe equipment operation.

Regenerating for prettier labels without source dimensions can repeat or worsen the safety defect. The model cannot determine a certified clearance by appearance alone.

OCR can transcribe the label but does not measure scale or prove that the drawn geometry meets engineering specifications. The numeric value requires independent dimensional validation.

Question 411

An image-generation API returns an encoded image payload rather than an HTTPS URL, but the client assumes every response contains a link. What should the app do?

Publish every generated image to an anonymous public bucket to simplify display.
Render the encoded text directly as an image URL without decoding it.
Handle the documented binary/base64 response variant, validate media type and store output securely if needed.
Generate a fake CDN URL using the model's response text.

Explanation

Correct Answer

Handle the actual response variant, decode bytes when necessary, verify media type and store or display through a protected output path. A successful generation is not necessarily a downloadable HTTPS link.

Incorrect Answers

Publishing generated images to an anonymous host may leak unreleased creative work or customer prompts. Storage/access should be governed by business needs.

Base64 or binary image output must be decoded or handled according to the documented media type. Treating encoded text as a URL produces an invalid resource.

A fabricated CDN URL cannot point to image data that was never uploaded. It creates a misleading success result.

Question 412

A designer wants five visually different concepts from one brief. A teammate claims increasing sampling diversity guarantees correct printed pricing in all five images. What is accurate?

Use a fixed random seed for every candidate and assume repeated rendering makes printed prices authoritative.
Sampling choices may change visual variety, but exact commercial text still needs independent verification.
Reduce generation variability and verify only the first candidate; subsequent images with the same prompt share its price accuracy.
Generate multiple variants and select the one with the clearest-looking price typography without comparing it to the current price catalog.

Explanation

Correct Answer

Image-generation settings can change output diversity but do not validate factual retail information. Every candidate that may be published must have its essential prices and claims verified against approved source data.

Incorrect Answers

A fixed seed may aid some reproducibility workflows, depending on the supported model, but it does not prove that generated pixel text matches a binding catalog price.

A lower-variance output can still contain erroneous numbers, and generation results are not guaranteed identical. Every selected candidate needs the same release controls for commercial text.

A plausible-looking numeric label is not evidence of the merchant’s current amount. The catalog remains authoritative, even when the generated typography looks sharp.

Question 413

A product-image editor keeps adding a second lamp when asked to change only the background. What refinement is most appropriate?

Add a negative prompt saying “no extra lamps,” but continue unrestricted whole-image editing and approve without visual comparison.
Increase whole-image variation strength so the background changes more dramatically, without preserving the foreground mask.
Compare only the generated image dimensions with the original and accept the edit whenever they match.
Clarify the edit scope and use supported reference/mask controls, then inspect the actual result for unwanted additions.

Explanation

Correct Answer

Use supported mask/reference editing controls, specify that product geometry and item count must remain fixed, then compare the output to the source and reject unintended additions.

Incorrect Answers

Negative prompting can guide generation but does not enforce which pixels may change. Without a controlled edit region and source comparison, the model may still add lamps or change the product.

Greater variation can change product geometry and item count as well as the background. It does not protect the previously approved foreground from unwanted alterations.

Equal width and height do not prove that the original foreground or number of lamps was preserved. Inspect the content of the edited and protected regions.

Question 414

An image generation prompt contains an unreleased product specification and an internal marketing brief. The app logs every prompt to a broadly shared analytics tool. What control is needed?

Restrict source and output access, minimize sensitive prompt details and enforce retention rules for logs.
Assume generated imagery removes confidentiality requirements from input briefs.
Mask only the final picture while preserving unrestricted prompt logs.
Move the full confidential prompts into a team-wide encrypted analytics workspace without narrowing staff access or setting retention limits.

Explanation

Correct Answer

Confidential briefs and unreleased specifications require restricted model input, output and logs. Minimize sensitive prompt content, enforce authorized access and retention, and avoid broadcasting raw internal text through analytics.

Incorrect Answers

Generated artwork does not automatically declassify the proprietary content supplied to make it. Input and intermediate traces remain governed company information.

Masking a final picture does nothing to stop analytics viewers from reading the unrestricted original brief in prompt logs.

Encryption protects storage media but cannot justify broadcasting unreleased product specifications to unnecessary users or retaining them indefinitely. Access scope and data minimization remain required.

Question 415

A user wants an image with a genuinely transparent background, but a deployment only supports opaque images. What should the app verify?

Judge transparency solely by whether the background appears white in a viewer.
Request "transparent" in text and assume every model produces an alpha channel.
Whether alpha/transparency is supported for the selected model/output format and validate the actual image channel.
Save an opaque JPEG with the filename suffix `.png` and claim transparency.

Explanation

Correct Answer

Confirm that the deployed model and selected output format support transparency, then inspect actual alpha values or channel metadata in the produced image.

Incorrect Answers

A white area in a viewer can be opaque RGB pixels rather than transparency. Visual appearance alone cannot confirm an alpha channel.

A “transparent” prompt cannot override a model/API that only produces opaque outputs. Capability must exist in the selected generation workflow.

Renaming opaque JPEG bytes to PNG neither changes the encoding nor creates transparency. A valid transparent asset needs actual alpha-supporting output.

Question 416

A company changes its image-generation model version before launching a regulated customer campaign. Select TWO release considerations.

Assume identical results because the new model belongs to the same product family.
Compare safety, brand fidelity and input/output compatibility on representative prompts and references.
Ignore all prior samples so reviewers cannot compare drift.
Review usage rights and deployment terms for the specific new model/version before distribution.
Publish without checking whether the new API still supports the required image output.

Explanation

Correct Answers

Run representative before-and-after prompts and reference inputs, assessing brand fidelity, safety, size and response-format compatibility. Record differences before distributing regulated campaign assets.

Check the specific version’s license, commercial-use terms, deployment restrictions and relevant asset rights. A model update can change conditions independently of image quality.

Incorrect Answers

Shared product family does not guarantee identical safety filters, image formats or visual quality. A revision may materially change rendered brand details.

Discarding prior samples removes the only baseline for identifying unwanted model drift and makes regressions hard to explain.

An API version may stop supporting required image-output shapes. Publishing without compatibility tests risks broken or missing campaign content.

Question 417

An existing Python integration uses the Azure Vision Image Analysis 4.0 SDK to caption an image from a URL. Which initialization is consistent with that SDK while the team plans migration before its announced retirement?

Create the supported `ImageAnalysisClient` with the Vision endpoint and approved credential, then request appropriate visual features.
Send the image URL to an unrelated text-only chat deployment and assume it ran image analysis.
Skip credentials because URL-based image analysis never requires service authentication.
Initialize the image analysis client with a speech-to-text language string as the endpoint.

Explanation

Correct Answer

Use ImageAnalysisClient with the Azure Vision endpoint and approved AzureKeyCredential or supported credential type, then request captioning or other documented visual features. Retain a migration plan for the service’s retirement schedule.

Incorrect Answers

A text-only model given a URL string does not perform a Vision 4.0 image analysis call or guarantee inspection of the actual image.

A public image URL is only an input locator and does not bypass authentication to the protected Vision resource.

A spoken-language code is not an HTTPS Vision endpoint and cannot identify the resource used by ImageAnalysisClient.

Question 418

A Vision client requests only an image caption but now needs exact serial digits printed on equipment. Which feature choice should be added or substituted?

Crop and upscale the serial-number region, then trust any digits suggested by the existing image-caption model without using text recognition.
Use an image-level caption of the equipment label and parse any serial-like digits appearing in the generated description.
Increase the caption's maximum length and treat any number it mentions as exact OCR.
An OCR/read-capable visual feature with character and source-region checks.

Explanation

Correct Answer

Use a Read/OCR-capable visual feature with source-region and character verification. Capture or crop high-resolution label evidence and flag unreadable characters before writing the serial.

Incorrect Answers

Cropping can improve input visibility, but a caption model is not obliged to reproduce exact alphanumeric characters. Read/OCR capability plus source-character verification is required for a trustworthy equipment serial.

A caption can identify a machine and describe its label without preserving every printed character. Serial-number evidence requires a text-reading capability and comparison against the actual visible source region.

Longer captions can mention plausible digits without reading them exactly. A narrative is not an OCR transcript.

Question 419

An image catalog client already calls an Azure Vision model. Its UI expects several independent labels for each uploaded product photo, but the client currently stores only one generated caption sentence. What implementation change matches the requirement?

Extract candidate nouns from the caption and assign the caption’s single confidence score to each noun as though it were a separately detected tag.
Run an image-level classifier for the dominant product category and assign that one label to every object present in the photo.
Request or parse the supported image-tagging output, retain individual tag labels with their confidence evidence, and evaluate them against the catalog taxonomy.
Ask a generative vision model for a comma-separated list of labels and treat them as native scored tags without checking the supported tag output.

Explanation

Correct Answer

Request supported image tag results rather than a scene caption, preserve each label and available confidence/source evidence, and map only valid tags to the approved product taxonomy.

Incorrect Answers

A caption is one scene description, not a list of independently scored categories. Reusing its score for noun fragments invents per-tag confidence and can mislabel the actual objects.

A dominant category does not yield the multiple independent labels the catalog needs. It can hide secondary items and cannot provide the appropriate per-tag evidence.

An unconstrained generated list does not carry the confidence and taxonomy contract of the image-tagging response. It can include inferred objects absent from the source image.

Question 420

An archival photo application needs a one-sentence description of an overall scene for search. It does not need individual object boxes. What output should it evaluate?

Image captioning with source-image quality checks and confidence/uncertainty handling.
Only a list of OCR-detected printed letters.
A newly generated stock photograph with a plausible scene.
Bounding boxes without any natural-language scene description.

Explanation

Correct Answer

Captioning summarizes the actual scene in a concise sentence. Validate factual coverage and unsupported-detail rates against the archival image before indexing.

Incorrect Answers

OCR extracts visible letters and cannot describe the scene when the archive needs an overall visual summary.

A newly generated stock image is not the historical source and cannot describe what the original photograph depicts.

Bounding boxes help locate objects but may not provide the requested coherent scene-level description, such as interactions and surroundings.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests

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)
Microsoft AI-901 Practice Test 43 - 10 Questions (Last Reviewed 02-Oct-2026)
Microsoft AI-901 Practice Test 44 - 10 Questions (Last Reviewed 01-Oct-2026)

Pass Your Microsoft AI-901 Exam Easy!

Question 410

Explanation

Correct Answer

Incorrect Answers

Question 411

Explanation

Correct Answer

Incorrect Answers

Question 412

Explanation

Correct Answer

Incorrect Answers

Question 413

Explanation

Correct Answer

Incorrect Answers

Question 414

Explanation

Correct Answer

Incorrect Answers

Question 415

Explanation

Correct Answer

Incorrect Answers

Question 416

Explanation

Correct Answers

Incorrect Answers

Question 417

Explanation

Correct Answer

Incorrect Answers

Question 418

Explanation

Correct Answer

Incorrect Answers

Question 419

Explanation

Correct Answer

Incorrect Answers

Question 420

Explanation

Correct Answer

Incorrect Answers

Microsoft Azure AI Fundamentals (AI-901) Practice Tests