Microsoft Certified: Azure AI Fundamentals (AI-900) — Questions and Answers

Question 1: What does the 'Accountability' principle in Microsoft's Responsible AI framework require?

All AI training data must be publicly disclosed
People and organizations must take responsibility for AI systems and their impact (Correct answer)
AI systems must be open-source
AI must always defer to human decisions

Correct answer: People and organizations must take responsibility for AI systems and their impact

The Accountability principle holds people and organizations responsible for the design, deployment, and outcomes of their AI systems.

Question 2: Which of these metrics is not applied to Custom Vision model training?

Precision
F1 Score (Correct answer)
Recall
Mean Average Precision

Correct answer: F1 Score

While F1 Score is a common metric in machine learning, Azure Custom Vision's portal and documentation primarily emphasize Precision, Recall, and Mean Average Precision (mAP) as the key performance indicators displayed for model training results. Therefore, F1 Score is not typically presented as a primary, directly 'applied' metric within the service's user interface compared to the others.

Question 3: Which Azure AI Vision feature can generate a dense description of all regions within an image?

Dense captioning (Correct answer)
Object tagging
Spatial analysis
Smart cropping

Correct answer: Dense captioning

Dense captioning generates individual natural language captions for multiple regions across an image, providing a richer description than a single overall caption.

Question 4: Which Azure service can analyze text in multiple languages and return the detected language with a confidence score?

Azure AI Speech
Azure AI Vision
Azure AI Language – Language Detection (Correct answer)
Azure AI Search

Correct answer: Azure AI Language – Language Detection

The Language Detection feature within Azure AI Language identifies the language of text input and provides an ISO language code along with a confidence score.

Question 5: A retailer wants to predict next month's sales figures based on historical data. Which type of machine learning task is this?

Clustering
Anomaly detection
Regression (Correct answer)
Classification

Correct answer: Regression

Regression predicts a continuous numerical value, such as a sales figure, based on input features.

Question 6: What is the purpose of a 'model card' in Responsible AI documentation?

A card game used to teach AI concepts to students
A structured document that describes a model's intended use, performance metrics, limitations, and ethical considerations (Correct answer)
A credit card used to pay for Azure AI services
A business card format for AI researchers to share contact information

Correct answer: A structured document that describes a model's intended use, performance metrics, limitations, and ethical considerations

A model card is a transparency artifact that summarizes what a model does, how it performs across different groups, its limitations, and the contexts in which it should or should not be used.

Question 7: A hospital wants to predict which patients are at high risk of readmission within 30 days of discharge. Which type of ML task is this?

Regression
Binary classification (Correct answer)
Object detection
Clustering

Correct answer: Binary classification

Predicting whether a patient will or will not be readmitted is a binary classification task with two possible outcome classes: readmitted or not readmitted.

Question 8: What NLP task involves identifying people, places, and organizations mentioned in text?

Language detection
Named Entity Recognition (NER) (Correct answer)
Sentiment analysis
Key phrase extraction

Correct answer: Named Entity Recognition (NER)

Named Entity Recognition (NER) is the NLP task that identifies and classifies entities such as people, locations, and organizations in text.

Question 9: Which Microsoft tool allows business users and subject-matter experts to create chatbots without writing code?

Azure Bot Service
Power Virtual Agents (Correct answer)
LUIS Studio
Bot Framework SDK

Correct answer: Power Virtual Agents

Power Virtual Agents is a low-code/no-code platform that enables non-developers to build and deploy bots using a guided graphical interface.

Question 10: Which Azure AI Language feature is the official successor to LUIS for building conversational natural language understanding models?

Conversational Language Understanding (CLU) (Correct answer)
Custom Named Entity Recognition
Sentiment Analysis
Custom Text Classification

Correct answer: Conversational Language Understanding (CLU)

Conversational Language Understanding (CLU) within Azure AI Language is the next-generation replacement for LUIS, offering intent recognition and entity extraction with an updated authoring experience.

Question 11: Which of the following doesn't include artificial intelligence in some way?

Face recognition system
LIDAR
Chatbots
DBMS (Correct answer)

Correct answer: DBMS

A Database Management System (DBMS) is software used for managing databases, including data storage, retrieval, and manipulation. While modern DBMS might incorporate AI for optimization or advanced analytics, a core DBMS itself is not inherently an AI system. In contrast, LIDAR, Chatbots, and Face recognition systems are technologies whose core functionalities are heavily reliant on or are direct applications of Artificial Intelligence.

Question 12: Which NLP concept describes the multiple meanings a single word can have depending on context?

Tokenization
Stemming
Lexical ambiguity (Correct answer)
Lemmatization

Correct answer: Lexical ambiguity

Lexical ambiguity refers to the phenomenon where a word has multiple possible meanings that must be resolved using surrounding context.

Question 13: Which metric is commonly used to evaluate object detection model performance in Azure Custom Vision?

Mean Average Precision (mAP) (Correct answer)
Word Error Rate (WER)
Root Mean Squared Error (RMSE)
Perplexity

Correct answer: Mean Average Precision (mAP)

Mean Average Precision (mAP) measures how accurately an object detection model identifies and localizes objects across all classes.

Question 14: What is the purpose of sentiment analysis in Azure AI Language?

Translate text between languages
Summarize long documents
Determine if text expresses positive, negative, or neutral opinion (Correct answer)
Extract named entities from text

Correct answer: Determine if text expresses positive, negative, or neutral opinion

Sentiment analysis evaluates text and assigns a sentiment label (positive, negative, or neutral) along with a confidence score.

Question 15: Which Azure AI service would best support adding real-time captions and transcriptions to a video conferencing application?

Azure AI Document Intelligence
Azure AI Speech (Correct answer)
Azure AI Vision
Azure AI Language

Correct answer: Azure AI Speech

Azure AI Speech provides real-time speech recognition that can generate live captions and transcriptions for audio/video streams.

Question 16: In Azure Bot Service, what is a 'channel'?

A communication pathway connecting the bot to users on platforms like Microsoft Teams or Slack (Correct answer)
A type of neural network layer used in bot training
An API endpoint used during bot registration
A database connection for storing conversation history

Correct answer: A communication pathway connecting the bot to users on platforms like Microsoft Teams or Slack

A channel is a connection between the bot and a communication platform (e.g., Teams, Slack, Web Chat), allowing users to interact with the bot on their preferred service.

Question 17: Which Azure tool helps teams detect and mitigate fairness issues in machine learning models?

Fairlearn (integrated with Azure Machine Learning) (Correct answer)
Azure Policy
Azure Monitor
Azure Blueprints

Correct answer: Fairlearn (integrated with Azure Machine Learning)

Fairlearn is an open-source toolkit integrated with Azure Machine Learning that helps assess and mitigate fairness-related harms in machine learning models.

Question 18: What Azure service analyzes video feeds to understand people's movements and spatial behavior?

Azure AI Video Indexer
Azure Media Services
Azure Digital Twins
Azure AI Vision Spatial Analysis (Correct answer)

Correct answer: Azure AI Vision Spatial Analysis

Azure AI Vision Spatial Analysis processes live video to detect people, track movement, and understand spatial interactions in physical spaces.

Question 19: Which Azure feature allows you to extract structured information from forms and invoices?

Azure AI Document Intelligence (Form Recognizer) (Correct answer)
Azure AI Speech
Azure AI Language key phrase extraction
Azure Translator

Correct answer: Azure AI Document Intelligence (Form Recognizer)

Azure AI Document Intelligence (formerly Form Recognizer) uses AI to extract structured data like fields and tables from documents and forms.

Question 20: What type of output does Azure Custom Vision's object detection model return?

Bounding box coordinates and class labels for each detected object (Correct answer)
A textual description of the image
A single class label for the image
A sentiment score for the scene

Correct answer: Bounding box coordinates and class labels for each detected object

Custom Vision object detection returns bounding box coordinates (left, top, width, height) along with class labels and confidence scores for each detected object.

Question 21: What is 'data poisoning' in the context of AI security?

Adding noise to model weights during training
Encrypting training data to prevent unauthorized access
Intentionally introducing malicious or misleading data into a training dataset to corrupt the model's behavior (Correct answer)
Oversampling minority classes to improve fairness

Correct answer: Intentionally introducing malicious or misleading data into a training dataset to corrupt the model's behavior

Data poisoning is an adversarial attack where malicious data is injected into a training set to manipulate the model's learned behavior in harmful ways.

Question 22: In the Bot Framework, what is a 'dialog'?

A LUIS feature for grouping related utterances together
A modular, reusable component that manages a specific part of the conversation flow (Correct answer)
A channel connector used to link the bot to communication platforms
A text response generated and returned by the bot to the user

Correct answer: A modular, reusable component that manages a specific part of the conversation flow

A dialog is a self-contained component in Bot Framework that encapsulates a specific conversational task (e.g., collecting a booking date), enabling reuse and composability.

Question 23: What is the function of the 'intent' in a Conversational Language Understanding (CLU) model?

A confidence threshold for model predictions
A translation target for multilingual bots
The predicted goal or action the user wants to perform (Correct answer)
A word-level label for named entities

Correct answer: The predicted goal or action the user wants to perform

In CLU, an intent represents the purpose or goal behind a user's utterance, such as 'BookFlight' or 'CheckWeather'.

Question 24: What does 'stemming' accomplish in NLP preprocessing?

Translates words to another language
Converts words to their root or base form by removing suffixes (Correct answer)
Splits text into individual characters
Identifies named entities in a sentence

Correct answer: Converts words to their root or base form by removing suffixes

Stemming reduces words to their base form (e.g., 'running' → 'run') by stripping affixes, helping normalize text for analysis.

Question 25: What does OCR stand for in the context of Azure AI Vision?

Online Content Recognition
Output Content Rendering
Object Classification Routine
Optical Character Recognition (Correct answer)

Correct answer: Optical Character Recognition

OCR stands for Optical Character Recognition, which is the ability to extract printed or handwritten text from images.

Question 26: In Azure AI Search, what is a 'skillset'?

A pricing bundle for AI Search API calls
A pipeline of cognitive skills that enrich indexed documents with AI-extracted information (Correct answer)
A list of developer certifications required to use the service
A set of search query templates stored in the index

Correct answer: A pipeline of cognitive skills that enrich indexed documents with AI-extracted information

A skillset in Azure AI Search is a collection of cognitive skills (like OCR, entity extraction, or language detection) applied during indexing to enrich documents.

Question 27: Which concept describes the practice of testing AI systems with diverse and adversarial inputs before deployment?

Red teaming (Correct answer)
Model compression
Transfer learning
Hyperparameter tuning

Correct answer: Red teaming

Red teaming involves probing an AI system with adversarial, edge-case, or unexpected inputs to identify failure modes and vulnerabilities before deployment.

Question 28: In a knowledge base built with Azure AI Language Custom Question Answering, what are the fundamental building blocks?

Intents and entities
Question-and-answer pairs (Correct answer)
Nodes and directed edges
Training datasets and validation sets

Correct answer: Question-and-answer pairs

A Custom Question Answering knowledge base is composed of QA pairs — each entry has one or more question phrasings matched to a specific answer.

Question 29: What does the Direct Line channel in Azure Bot Service allow developers to do?

Enable voice-only communication through Azure AI Speech
Enable direct bot-to-bot communication without a user
Connect custom client applications to a bot using a REST API (Correct answer)
Link the bot exclusively to Azure internal services

Correct answer: Connect custom client applications to a bot using a REST API

The Direct Line channel provides a REST API and WebSocket interface that lets custom-built client apps (web, mobile, IoT) communicate directly with an Azure bot.

Question 30: You are developing an app with AI. Make sure the app adheres to the responsible AI guidelines. Which two guidelines ought to you abide by?

Establish a risk governance committee that includes members of the legal team, members of the risk management team, and a privacy officer (Correct answer)
Implement an Agile software development methodology
Prevent the disclosure of the use of AI-based algorithms for automated decision making
Implement a process of AI model validation as part of the software review process (Correct answer)

Correct answer: Establish a risk governance committee that includes members of the legal team, members of the risk management team, and a privacy officer

Adhering to responsible AI guidelines involves both technical and organizational measures. Implementing a robust AI model validation process ensures the model's performance, fairness, and reliability are continuously checked. Additionally, establishing a risk governance committee with diverse expertise (legal, risk management, privacy) provides essential oversight and ensures ethical considerations are addressed throughout the AI system's lifecycle, promoting accountability and responsible development.

Question 31: What is a 'confusion matrix' used for in machine learning evaluation?

Comparing training vs. validation loss curves
Visualizing neural network architecture
Plotting feature importance scores
Summarizing correct and incorrect predictions across classes for classification models (Correct answer)

Correct answer: Summarizing correct and incorrect predictions across classes for classification models

A confusion matrix shows true positives, true negatives, false positives, and false negatives, providing a complete picture of a classifier's performance per class.

Question 32: Which principle of responsible AI ensures that AI systems can be examined and their decisions understood?

Reliability
Transparency (Correct answer)
Privacy
Fairness

Correct answer: Transparency

Transparency means AI systems and their decision-making processes should be understandable and explainable to developers, operators, and affected users.

Question 33: Which component of an Azure ML workspace stores the versioned datasets, models, and experiment run history?

Inference cluster
Compute instance
Environment definition
Datastore and registry artifacts (Correct answer)

Correct answer: Datastore and registry artifacts

Datastores link to Azure storage accounts to hold datasets, while the workspace registry tracks versioned models and experiment artifacts for reproducibility.

Question 34: What is the 'Read API' in Azure AI Vision used for?

Processing audio read requests from users
Reading JSON responses from other Azure APIs
Extracting printed and handwritten text from images and PDFs at scale (Correct answer)
Reading Azure resource configuration settings

Correct answer: Extracting printed and handwritten text from images and PDFs at scale

The Read API in Azure AI Vision is an OCR engine optimized for extracting large amounts of printed and handwritten text from images and multi-page PDFs.

Question 35: What is a 'tag' in the context of Azure AI Vision image analysis?

An error code returned when image analysis fails
A metadata label attached to an Azure resource
A keyword describing objects, actions, or concepts visible in an image (Correct answer)
A bounding box coordinate for object detection

Correct answer: A keyword describing objects, actions, or concepts visible in an image

Azure AI Vision returns tags — descriptive keywords with confidence scores that identify recognizable objects, scenes, and actions present in an image.

Question 36: Which Azure service provides pre-built APIs for image analysis, object detection, and OCR?

Azure AI Vision (Correct answer)
Azure AI Language
Azure Bot Service
Azure AI Speech

Correct answer: Azure AI Vision

Azure AI Vision (part of Azure Cognitive Services) offers pre-built APIs for analyzing images, detecting objects, reading text, and more.

Question 37: Which of the following best describes 'natural language generation' (NLG)?

Translating text from one language to another
Converting structured data or prompts into human-readable text (Correct answer)
Analyzing the meaning and intent behind text
Converting spoken audio into written text

Correct answer: Converting structured data or prompts into human-readable text

NLG is the AI capability of producing coherent, human-readable text from structured data, templates, or learned patterns.

Question 38: In Azure Custom Vision, what is the minimum recommended number of images per tag for a good starting model?

5 images
15 images (Correct answer)
50 images
500 images

Correct answer: 15 images

Azure Custom Vision recommends at least 15 images per tag as a starting point, though more images generally lead to better model accuracy.

Question 39: Which Microsoft tool provides a dashboard for assessing error analysis, fairness, and model interpretability in Azure Machine Learning?

Azure Advisor
Responsible AI Dashboard in Azure Machine Learning (Correct answer)
Azure Cost Management
Azure Security Center

Correct answer: Responsible AI Dashboard in Azure Machine Learning

The Responsible AI Dashboard in Azure Machine Learning consolidates tools for error analysis, fairness assessment, model explanations, and data exploration in one interface.

Question 40: Which responsible AI principle is most directly concerned with protecting personal information used by AI systems?

Transparency
Privacy and security (Correct answer)
Fairness
Reliability and safety

Correct answer: Privacy and security

The privacy and security principle requires that AI systems protect personal data and give users control over their information.

Microsoft Certified: Azure AI Fundamentals (AI-900)

This exam validates foundational knowledge of machine learning and artificial intelligence (AI) concepts and related Microsoft Azure AI services.

Questions

60 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds