Microsoft AI-901 Practice Exam

• A. Fairness Correct answer
• B. Inclusiveness
• C. Transparency
• D. Privacy and security

Option A, Fairness, is correct because the fairness principle requires an AI system to treat all people and groups equitably, rather than affecting similar groups differently. Inclusiveness is about engaging people of all abilities, transparency is about understanding how the system works, and privacy and security is about protecting data.

• A. Accountability
• B. Reliability and safety
• C. Transparency
• D. Fairness Correct answer

Option D, Fairness, is correct because approving equally qualified applicants at different rates by gender is unequal treatment, which is exactly what fairness prohibits. Accountability, reliability and safety, and transparency concern ownership, dependable operation, and explainability, not equitable outcomes.

• A. Historical bias in the training data Correct answer
• B. Normalization of numeric features before training
• C. Anonymization of personal data before training
• D. Splitting data into training and test sets

Option A is correct because models learn patterns from their training data, so historical bias in that data is reproduced as unfair outcomes. Normalizing features, anonymizing personal data, and splitting data into training and test sets are standard practices that do not by themselves introduce bias against groups.

• A. An access credential that authenticates requests to the model endpoint
• B. A single numeric weight stored inside the neural network
• C. A labelled example drawn from the training dataset
• D. A word or word fragment the model reads and generates Correct answer

Option D is correct because a token is the unit of text, a whole word or a fragment of one, that the model reads as input and produces as output. Access credentials, network weights and labelled training examples are different concepts, even though API keys are sometimes also called tokens.

• A. Output sampling layer
• B. Embedding layer
• C. Tokenizer
• D. Self-attention Correct answer

Option D is correct because self-attention computes how much each token should attend to every other token in the sequence, which is the core mechanism of the transformer. The tokenizer only splits text, the embedding layer maps tokens to vectors, and the sampling layer picks the next token from the output probabilities.

• A. Prompt a chat model with a longer system message describing every article
• B. Compare embedding vectors of the question and each article Correct answer
• C. Raise the temperature so the model explores related wording
• D. Tokenize both texts and rank articles by the number of shared tokens

Option B is correct because embeddings turn text into vectors that capture meaning, so a question and an article with no shared keywords can still be close in vector space. Counting shared tokens depends on overlapping wording, temperature only changes sampling randomness, and a longer system message does not perform a comparison across many articles.

• A. Conversational AI
• B. Forecasting
• C. Anomaly detection Correct answer
• D. Information extraction

Option C is correct because anomaly detection finds data points, such as card transactions, that deviate sharply from an established normal pattern. Forecasting predicts future values, conversational AI handles dialogue, and information extraction pulls structured data out of documents, so none of them flag unusual spending.

• A. Anomaly detection
• B. Generative AI Correct answer
• C. Information extraction
• D. Computer vision

Option B is correct because generative AI creates new content such as text, code or images from a natural-language prompt. Anomaly detection, information extraction and computer vision analyze existing data rather than producing new content from a prompt.

• A. Tagging holiday photos by their content Correct answer
• B. Turning a script into spoken audio
• C. Predicting delivery times from past trips
• D. Translating a support chat into Portuguese
• E. Spotting hard hats on site camera frames Correct answer

Options A and E are correct because tagging photos by content and spotting hard hats in camera frames both extract meaning from images, which is a computer vision workload. Text-to-speech is speech, delivery-time prediction is machine learning on tabular data, and translating chat is a natural language workload.

• A. Tool message
• B. System message Correct answer
• C. Assistant message
• D. User message

Option B is correct because the system message sets the assistant's persona, tone and ground rules for the whole conversation. User messages carry the end user's input, assistant messages hold earlier model replies, and tool messages return tool call results.

• A. Extend the model's context window
• B. Lower the cost of each request
• C. Show the desired output format Correct answer
• D. Update the model's weights

Option C is correct because few-shot examples show the model sample inputs and outputs, so it imitates the desired format and style. They do not change the context window size, reduce cost (they add tokens), or update the model's weights, which only training or fine-tuning does.

• A. Supplying sample replies for few-shot prompts Correct answer
• B. Replaying earlier replies as history Correct answer
• C. Defining the assistant's persona
• D. Setting the temperature for the request
• E. Passing the user's latest question

Options A and B are correct because assistant messages can carry sample replies in few-shot prompts and replay the model's earlier replies as conversation history. The persona belongs in the system message, temperature is a request parameter, and the user's latest question goes in a user message.

• A. Sentiment analysis Correct answer
• B. Key phrase extraction
• C. Entity linking
• D. Language detection

Option A is correct because sentiment analysis scores text as positive, neutral or negative. Key phrase extraction pulls out talking points, entity linking ties names to knowledge-base entries, and language detection only identifies the language.

• A. azure-ai-documentintelligence
• B. azure-ai-textanalytics Correct answer
• C. azure-cognitiveservices-speech
• D. azure-ai-vision-imageanalysis

Option B is correct because TextAnalyticsClient ships in the azure-ai-textanalytics package. The Document Intelligence, Speech and Vision packages expose different clients for forms, speech and image analysis.

• A. redacted_text
• B. url Correct answer
• C. subcategory
• D. category
• E. data_source Correct answer

Options B and E are correct because each linked entity carries a url to its knowledge-base page and a data_source naming that source, such as Wikipedia. category and subcategory belong to recognized entities from recognize_entities, and redacted_text belongs to PII detection.

• A. Text embedding model
• B. Multimodal chat model Correct answer
• C. Document translation model
• D. Speech-to-text model

Option B is correct because a multimodal chat model accepts both image and text content in one prompt. Embedding (A), document translation (C) and speech-to-text (D) models work on a single data type and do not interpret images.

• A. As a function definition in the tools array
• B. As a query-string value on the endpoint URL
• C. As an image_url part of the user message Correct answer
• D. As extra text in the system message

Option C is correct because chat completions take the image as an image_url content part inside the user message, next to the text part. Tool definitions (A), a URL query string (B) and system message text (D) are not ways to pass an image.

• A. Read the printed words on a street sign Correct answer
• B. Update its own weights from the photo
• C. Name the private person shown from the face
• D. Return a retouched version of the photo
• E. Write a caption describing the scene Correct answer

Options A and E are correct because a multimodal model such as gpt-4o can read printed text in an image and describe a scene in words. Option B is wrong because inference never changes model weights, option C is wrong because responsible-use rules block identifying private people from their faces, and option D is wrong because a chat model returns text, not an edited image.

• A. prebuilt-documentSearch
• B. prebuilt-invoice Correct answer
• C. prebuilt-document
• D. prebuilt-receipt

Option B is correct because prebuilt-invoice is the domain-specific analyzer whose schema already covers vendor name, invoice date and totals. prebuilt-documentSearch targets RAG ingestion, prebuilt-document is a base analyzer with no tuned fields, and prebuilt-receipt is for sales receipts.

• A. A custom analyzer with its own field schema Correct answer
• B. A fine-tuned model trained on sample leases
• C. A search index over the lease files
• D. A prebuilt analyzer with extra fields added to it

Option A is correct because when no prebuilt analyzer fits, you define a custom analyzer with your own field schema (optionally inheriting from a base analyzer such as prebuilt-document). Fine-tuning a model and building a search index do not produce structured field extraction, and a prebuilt analyzer cannot simply have fields added without creating a custom analyzer from it.