I created these AI-900 questionnaires to help people study for their certification, especially my colleagues. I am proudly a student from Simplon.
This questionnaire is based on the book "Azure AI Fundamentals (AI-900) Study Guide." I hope it helps you on your learning journey!
1. Which AI workload is designed to automatically detect and filter harmful or inappropriate content online?
- a. Knowledge mining
- b. Generative AI
- c. Content moderation
- d. Document intelligence
Show Answer
✅ Correct Answer: C
Why it's correct: Content moderation is specifically designed to detect and filter harmful or inappropriate content online.
Why the others are wrong: * Knowledge mining extracts insights from large datasets. * Generative AI produces new content. * Document intelligence focuses on data extraction from documents.
- a. To recommend content based on user preferences
- b. To generate new images
- c. To detect harmful content
- d. To analyze human language
Show Answer
✅ Correct Answer: A
Why it's correct: AI personalization aims to recommend content that aligns with a user’s preferences and behavior to enhance their experience.
Why the others are wrong: * Choice B is a function of Generative AI. * Choice C falls under Content Moderation. * Choice D is a core part of Natural Language Processing.
- a. Natural language processing (NLP)
- b. Content moderation
- c. Computer vision
- d. Document intelligence
Show Answer
✅ Correct Answer: A
Why it's correct: NLP is the workload dedicated to analyzing, interpreting, and generating human language.
Why the others are wrong: * Content moderation identifies inappropriate material, it doesn't "understand" language. * Computer vision deals with images and videos. * Document intelligence extracts info but isn't specialized in language analysis itself.
- a. NLP
- b. Knowledge mining
- c. Computer vision
- d. Generative AI
Show Answer
✅ Correct Answer: C
Why it's correct: Computer vision uses OCR to "see" and extract text from visual files.
Why the others are wrong: * NLP focuses on text analysis, not image content. * Knowledge mining derives insights from data but not via OCR. * Generative AI creates new content rather than extracting existing text.
- a. Document intelligence
- b. Knowledge mining
- c. Computer vision
- d. NLP
Show Answer
✅ Correct Answer: B
Why it's correct: Knowledge mining is used specifically to uncover hidden insights across vast amounts of unstructured data.
Why the others are wrong: * Document intelligence handles structured documents (like forms). * Computer vision handles visuals but not generalized insights. * NLP is for language tasks, not broad data mining.
- a. NLP
- b. Generative AI
- c. Knowledge mining
- d. Document intelligence
Show Answer
✅ Correct Answer: B
Why it's correct: Generative AI is designed to create entirely new, original material.
Why the others are wrong: * NLP processes language but doesn't focus on producing original multimedia. * Knowledge mining and Document intelligence work with *existing* data, they don't generate *new* data.
- a. Analyze Text API
- b. Analyze Image API
- c. Custom Categories API
- d. Moderate Text API
Show Answer
✅ Correct Answer: A
Why it's correct: The Analyze Text API is the specific tool within Content Safety for scanning text.
Why the others are wrong: * Analyze Image API scans pictures. * Custom Categories API defines groups but doesn't scan the text itself. * Moderate Text API is a broader term, but "Analyze Text" is the specific API name.
- a. Fairness
- b. Transparency
- c. Inclusiveness
- d. Accountability
Show Answer
✅ Correct Answer: A
Why it's correct: Fairness ensures that the system provides equal treatment regardless of demographic groups.
Why the others are wrong: * Transparency makes the AI understandable. * Inclusiveness makes AI accessible. * Accountability ensures developers take responsibility for the AI's behavior.
- a. NLP
- b. Knowledge mining
- c. Computer vision
- d. Generative AI
Show Answer
✅ Correct Answer: C
Why it's correct: Computer vision is the primary field for interpreting visual data.
Why the others are wrong: * NLP is for language. * Knowledge mining is for data insights. * Generative AI is for creating content, not necessarily analyzing existing visual media.
10. Which AI workload is responsible for processing large amounts of documents and extracting key information from them?
- a. NLP
- b. Generative AI
- c. Content moderation
- d. Document intelligence
Show Answer
✅ Correct Answer: D
Why it's correct: Document Intelligence (formerly Form Recognizer) is built specifically for processing large volumes of documents.
Why the others are wrong: * NLP understands language but doesn't specialize in document structure. * Generative AI creates content. * Content moderation filters material.
This is perfect for a GitHub README. Chapter 4 covers the core "Data Science" part of the AI-900, which is often the hardest part for students. Including the "Why the others are wrong" section will be extremely helpful for your colleagues to understand the difference between Regression, Classification, and Clustering.
Copy and paste this block into your README.md in VS Code below your Chapter 3 section:
Markdown
- a. To categorize data into distinct classes
- b. To cluster similar data points
- c. To predict a numerical outcome based on variables
- d. To analyze images and videos
Show Answer
✅ Correct Answer: C
Why it's correct: The primary purpose of regression is to predict a numerical/continuous value (like price or temperature).
Why the others are wrong: * Choice A describes Classification. * Choice B describes Clustering. * Choice D refers to Computer Vision/Deep Learning tasks.
- a. K-means clustering
- b. Predicting house prices based on features
- c. Segmenting customers based on purchase history
- d. Identifying anomalies in financial transactions
Show Answer
✅ Correct Answer: B
Why it's correct: Predicting house prices uses labeled data (known past prices) to train the model, which is the definition of supervised learning.
Why the others are wrong: * K-means and customer segmentation are unsupervised (they find patterns in unlabeled data). * Anomaly detection is often unsupervised or semi-supervised.
3. In binary classification, which algorithm is commonly used to predict probabilities between two classes?
- a. Linear regression
- b. Logistic regression
- c. Decision trees
- d. K-means
Show Answer
✅ Correct Answer: B
Why it's correct: Despite the name, Logistic Regression is a classification algorithm that outputs a probability (0 to 1) to put data into one of two categories.
Why the others are wrong: * Linear regression predicts continuous numbers, not class probabilities. * Decision trees can classify but don't inherently focus on probability scores. * K-means is for clustering, not classification.
- a. The model’s ability to distinguish between classes
- b. The average of errors in predictions
- c. A balance between precision and recall
- d. The total accuracy of the model
Show Answer
✅ Correct Answer: C
Why it's correct: The F1 score is the "harmonic mean" of precision and recall, providing a single metric that balances both.
Why the others are wrong: * Choice A is usually measured by AUC (Area Under Curve). * Choice B describes error metrics like MSE or MAE used in regression. * Choice D (Accuracy) is a different metric that can be misleading if classes are unbalanced.
- a. Training
- b. Inferencing
- c. Validation
- d. Data preparation
Show Answer
✅ Correct Answer: B
Why it's correct: Inferencing is the "live" phase where a trained model processes new data to provide an answer/prediction.
Why the others are wrong: * Training is the learning phase. * Validation is for checking performance during training. * Data preparation happens before the model even exists.
- a. Supervised
- b. Semisupervised
- c. Reinforcement
- d. Unsupervised
Show Answer
✅ Correct Answer: D
Why it's correct: K-means clustering works on unlabeled data to find natural groups, which is a core unsupervised learning task.
Why the others are wrong: * Supervised requires labels. * Reinforcement uses a reward/penalty system.
- a. Mean squared error (MSE)
- b. Coefficient of determination (R²)
- c. Root mean squared error (RMSE)
- d. Mean absolute error (MAE)
Show Answer
✅ Correct Answer: B
Why it's correct: R² (R-Squared) tells you what percentage of the data's variance is explained by the model (closer to 1.0 is better).
Why the others are wrong: * MSE, RMSE, and MAE all measure error (the distance between prediction and reality), not the variance explanation.
8. Which approach is used to address missing data by estimating based on patterns in available data?
- a. Mean imputation
- b. Predictive imputation
- c. Removal of incomplete data
- d. Data normalization
Show Answer
✅ Correct Answer: B
Why it's correct: Predictive imputation uses other features in the dataset to "guess" or predict what the missing value should be.
Why the others are wrong: * Mean imputation just uses the average (less accurate). * Removal deletes the data entirely, which can cause bias. * Normalization scales data (e.g., 0 to 1) but doesn't fill gaps.
- a. To identify hidden patterns in unlabeled data
- b. To predict numerical outcomes
- c. To assign data points to predefined categories
- d. To generate new content
Show Answer
✅ Correct Answer: C
Why it's correct: Classification is about putting data into buckets or categories (e.g., Spam vs. Not Spam).
Why the others are wrong: * Choice A is the goal of Clustering. * Choice B is the goal of Regression. * Choice D is the goal of Generative AI.
- a. Regression
- b. Classification
- c. Clustering
- d. Deep learning (DL)
Show Answer
✅ Correct Answer: C
Why it's correct: Clustering is the specific technique used to group unlabeled data points based on similarities.
Why the others are wrong: * Regression and Classification are supervised techniques (they need labels). * Deep learning is a broad field that can do clustering, but Clustering is the fundamental technique name.
- a. Classification
- b. Regression
- c. Clustering
- d. Deep learning (DL)
Show Answer
✅ Correct Answer: B
Why it's correct: Regression is the standard technique for predicting continuous numerical values (like sales or price).
Why the others are wrong: * Classification assigns data to categories. * Clustering groups similar data without making predictions. * Deep learning can do many things, but "Regression" is the specific name for numerical prediction.
2. Which metric in regression analysis measures the average error in predictions, regardless of positive or negative deviation?
- a. Mean squared error (MSE)
- b. Coefficient of determination (R²)
- c. Mean absolute error (MAE)
- d. Root mean squared error (RMSE)
Show Answer
✅ Correct Answer: C
Why it's correct: Mean Absolute Error (MAE) treats all errors equally by taking the absolute value, providing a straightforward average of how far off the predictions were.
Why the others are wrong: * MSE squares the errors (making large errors look much worse). * R² measures variance explanation, not error size. * RMSE is the square root of MSE; it still emphasizes larger errors more than MAE does.
3. In binary classification, what algorithm is commonly used to estimate the probability of a binary outcome?
- a. K-means clustering
- b. Logistic regression
- c. Linear regression
- d. Decision tree
Show Answer
✅ Correct Answer: B
Why it's correct: Logistic regression outputs a value between 0 and 1, which represents the probability of a piece of data belonging to a specific class.
Why the others are wrong: * K-means is for clustering. * Linear regression is for continuous numbers. * Decision trees classify but don't inherently provide a probability score like logistic regression.
- a. Reinforcement learning
- b. Supervised learning
- c. Unsupervised learning
- d. Genetic algorithms
Show Answer
✅ Correct Answer: B
Why it's correct: AutoML focuses on automating Classification, Regression, and Time-series forecasting, all of which are Supervised Learning tasks.
Why the others are wrong: It does not currently focus on Reinforcement or Unsupervised learning (like clustering).
5. In a multiclass classification problem, which approach builds a binary classifier for each class?
- a. K-means clustering
- b. One-vs-rest (OVR)
- c. Multinomial algorithm
- d. Logistic regression
Show Answer
✅ Correct Answer: B
Why it's correct: One-vs-Rest (OVR) creates a separate "Yes/No" classifier for every single category (e.g., Apple vs. Everything else, Orange vs. Everything else).
Why the others are wrong: * K-means is clustering. * Multinomial handles all classes in one model rather than separate binary ones.
6. Which Azure Machine Learning feature helps automate the process of trying multiple algorithms to find the best model?
- a. Custom script execution
- b. AutoML
- c. Dataset storage
- d. Deployment pipelines
Show Answer
✅ Correct Answer: B
Why it's correct: AutoML is designed to save time by automatically testing different algorithms and hyperparameters to see which one performs best.
Why the others are wrong: * Custom scripts are manual. * Dataset storage is just for data management. * Pipelines are for workflow automation, not algorithm selection.
7. What evaluation metric indicates the proportion of correctly identified positives in a classification model?
- a. Accuracy
- b. Precision
- c. Recall
- d. F1 score
Show Answer
✅ Correct Answer: C
Why it's correct: Recall (Sensitivity) measures how many of the actual positives the model successfully found.
Why the others are wrong: * Accuracy counts overall correctness. * Precision measures how many predicted positives were actually correct. * F1 Score is a balance of both.
8. Which ML technique is suitable for grouping data points based on similarities without prior labels?
- a. Supervised learning
- b. Classification
- c. Clustering
- d. Regression
Show Answer
✅ Correct Answer: C
Why it's correct: Clustering finds natural groupings in data when you don't have existing labels.
Why the others are wrong: Supervised learning (A), Classification (B), and Regression (D) all require labeled data.
9. In regression analysis, what term describes the average of squared differences between predicted and actual values?
- a. Mean absolute error (MAE)
- b. Root mean squared error (RMSE)
- c. Mean squared error (MSE)
- d. R² score
Show Answer
✅ Correct Answer: C
Why it's correct: MSE specifically squares the differences before averaging them.
Why the others are wrong: * MAE uses absolute values, not squares. * RMSE is the square root of the MSE.
- a. Support vector machine (SVM)
- b. Decision tree
- c. Neural network
- d. K-nearest neighbors
Show Answer
✅ Correct Answer: C
Why it's correct: Neural Networks are the foundation of Deep Learning, consisting of input, hidden, and output layers.
Why the others are wrong: SVM, Decision Trees, and KNN are "traditional" ML algorithms and do not use a layered neural architecture.
1. Which of the following is a fundamental concept in computer vision involving dividing an image into a grid of colored points?
- a. Filters
- b. Pixels
- c. Neural networks
- d. Labels
Show Answer
✅ Correct Answer: B
Why it's correct: Pixels are the fundamental building blocks of digital images, forming the grid of individual colored points that the computer "sees."
Why the others are wrong: * Filters modify pixels but aren't the structure itself. * Neural networks analyze the data within pixels. * Labels are categories assigned to the image.
- a. OCR
- b. Bounding boxes
- c. Facial detection
- d. Image classification
Show Answer
✅ Correct Answer: B
Why it's correct: Bounding boxes are the rectangular coordinates used to outline and locate specific objects in an image.
Why the others are wrong: * OCR reads text. * Facial detection only finds faces. * Image classification tells you what is in an image but not where it is.
- a. To detect color inversion
- b. To resize images
- c. To identify patterns in pixel data
- d. To assign labels
Show Answer
✅ Correct Answer: C
Why it's correct: Convolution is a feature extraction step used to identify patterns like edges, textures, and shapes within the pixel data.
Why the others are wrong: It isn't used for resizing or color inversion, and it happens before the final labeling step.
4. In Microsoft Azure’s computer vision tools, which service primarily handles tasks like object detection and facial analysis?
- a. Azure Cognitive Search Service
- b. Azure AI Vision
- c. Azure Machine Learning
- d. Azure Kubernetes Service
Show Answer
✅ Correct Answer: B
Why it's correct: Azure AI Vision (formerly Computer Vision) is the specific suite of tools for image analysis, OCR, and spatial analysis.
Why the others are wrong: * Search is for indexing data. * Machine Learning is a general platform. * Kubernetes is for managing containers.
5. What type of neural network is most commonly used for computer vision tasks like image classification?
- a. Recurrent neural network (RNN)
- b. Convolutional neural network (CNN)
- c. Generative adversarial network (GAN)
- d. Transformer
Show Answer
✅ Correct Answer: B
Why it's correct: CNNs are the industry standard for vision because they use convolutional layers to extract visual features effectively.
Why the others are wrong: * RNNs are for sequential data (like text/speech). * GANs are for generating new images. * Transformers are mostly for NLP (though "Vision Transformers" exist, CNNs are the primary answer for AI-900).
- a. Convolutional layers
- b. Fully connected layers
- c. Pooling layers
- d. Activation functions
Show Answer
✅ Correct Answer: C
Why it's correct: Pooling layers (like Max Pooling) downsample the image data, reducing the dimensions while keeping the most important information.
Why the others are wrong: Convolutional layers extract features; Fully connected layers connect neurons for the final decision.
- a. Image classification
- b. OCR
- c. Facial detection
- d. Object detection
Show Answer
✅ Correct Answer: B
Why it's correct: Optical Character Recognition (OCR) is the specific technology used to turn text images into machine-readable text data.
Why the others are wrong: Classification, Facial detection, and Object detection do not "read" or interpret characters.
- a. Lack of color accuracy
- b. Privacy and consent issues
- c. High computational costs
- d. Limited dataset availability
Show Answer
✅ Correct Answer: B
Why it's correct: Privacy and consent are the biggest ethical hurdles, as facial data is sensitive biometric information.
Why the others are wrong: Color accuracy and costs are technical issues, not ethical ones.
9. In Azure’s AI Vision Studio, what data accompanies object detection to indicate the model’s confidence?
- a. Pixel count
- b. Confidence score
- c. Bounding box color
- d. File type
Show Answer
✅ Correct Answer: B
Why it's correct: A confidence score (usually 0 to 1 or 0% to 100%) tells you how certain the AI is that its prediction is correct.
Why the others are wrong: Bounding box colors are just for visual display; pixel count and file type don't relate to accuracy.
- a. Identifying edge patterns
- b. Reducing image size
- c. Assigning probabilities to predictions
- d. Adding color to images
Show Answer
✅ Correct Answer: C
Why it's correct: Activation functions (like Softmax at the end) help map the network's internal values to a set of probabilities for each possible category.
Why the others are wrong: Edges are found by convolutional layers, and size is reduced by pooling layers.
1. Which of the following Azure AI Services enables instant translation of text into multiple languages?
- a. Azure AI Language
- b. Azure AI Speech
- c. Azure AI Translator
- d. Azure AI Sentiment
Show Answer
✅ Correct Answer: C
Why it's correct: Azure AI Translator is the specific service designed for text-to-text translation across dozens of languages.
Why the others are wrong: * Azure AI Language is for analyzing text (sentiment, entities). * Azure AI Speech is for converting voice to text (or vice versa). * Sentiment is a feature of AI Language, not a translation service.
- a. To analyze sentiment
- b. To detect language
- c. To identify and categorize entities
- d. To extract key phrases
Show Answer
✅ Correct Answer: C
Why it's correct: NER is used to find and label "entities" like people, places, dates, and organizations within a body of text.
Why the others are wrong: * Sentiment analysis checks for "Positive/Negative" tones. * Language detection identifies the language (e.g., French vs. English). * Key phrase extraction finds the main talking points but doesn't categorize them as specific entities.
- a. Text-to-speech
- b. Summarization
- c. Real-time speech-to-text
- d. Custom Translator
Show Answer
✅ Correct Answer: C
Why it's correct: Real-time speech-to-text allows you to transcribe audio streams (like a live meeting) into text instantly.
Why the others are wrong: * Text-to-speech goes the opposite way (writing to voice). * Summarization condenses text. * Custom Translator is for specialized text translation.
- a. Assigning a unique identifier to each entity
- b. Converting text to speech
- c. Breaking text into individual words or phrases
- d. Detecting language in text
Show Answer
✅ Correct Answer: C
Why it's correct: Tokenization is the process of splitting a sentence into smaller units (tokens), such as words or sub-words, so the AI can process them.
Why the others are wrong: It doesn't involve identifiers, voice conversion, or language detection.
- a. It categorizes entities in text.
- b. It highlights main concepts or themes.
- c. It detects the language of a document.
- d. It analyzes sentiment in text.
Show Answer
✅ Correct Answer: B
Why it's correct: Key phrase extraction quickly identifies the most important points or "talking points" in a document.
Why the others are wrong: * NER (Choice A) handles categories. * Language detection (Choice C) handles languages. * Sentiment analysis (Choice D) handles emotions.
6. Which NLP feature in Azure would be most suitable for identifying sensitive information like Social Security numbers?
- a. Language detection
- b. Key phrase extraction
- c. PII detection
- d. Sentiment analysis
Show Answer
✅ Correct Answer: C
Why it's correct: PII (Personally Identifiable Information) detection is designed to find and redact sensitive data like SSNs, emails, and phone numbers.
Why the others are wrong: None of the other features are built to recognize or protect private, sensitive data types.
7. Which term refers to removing common words that don’t add meaning, like "the" and "an", during NLP processing?
- a. Lemmatization
- b. Stop-word removal
- c. Tokenization
- d. Frequency analysis
Show Answer
✅ Correct Answer: B
Why it's correct: Stop-word removal cleans the text of common words that don't help the AI understand the core meaning of a sentence.
Why the others are wrong: * Lemmatization reduces words to their root (e.g., "running" to "run"). * Tokenization splits the text. * Frequency analysis counts how often words appear.
8. In the Azure AI Language Studio, which feature would you use to automatically link an entity like “Paris” to a specific reference?
- a. Entity recognition
- b. Entity linking
- c. Language detection
- d. Summarization
Show Answer
✅ Correct Answer: B
Why it's correct: Entity linking provides context by connecting a word to a specific knowledge base entry (e.g., distinguishing between Paris, France and Paris Hilton).
Why the others are wrong: Entity recognition finds the word, but "linking" provides the specific identity/reference.
- a. To translate text
- b. To provide quick, synchronous transcription of audio
- c. To detect sentiment in audio
- d. To convert text to lifelike speech
Show Answer
✅ Correct Answer: B
Why it's correct: The Fast Transcription API is optimized for speed, providing rapid transcription for audio files.
Why the others are wrong: Translation is handled by Translator; lifelike speech is handled by Text-to-Speech.
- a. Text-to-speech
- b. Summarization
- c. Language detection
- d. Entity recognition
Show Answer
✅ Correct Answer: B
Why it's correct: Summarization uses AI to condense long documents into short, meaningful summaries.
Why the others are wrong: Text-to-speech makes audio; detection and recognition identify parts of the text but don't condense it.
1. Which feature of generative AI on Azure allows for generating unique images based on text prompts?
- a. Semantic search
- b. DALL-E
- c. Content moderation
- d. Lifelike dialogue creation
Show Answer
✅ Correct Answer: B
Why it's correct: DALL-E is the specific generative AI model integrated into Azure OpenAI for creating original images from natural language descriptions.
Why the others are wrong: * Semantic search retrieves existing info. * Content moderation filters safe/unsafe content. * Dialogue creation is a text-based task (GPT).
- a. To identify harmful content
- b. To translate languages
- c. To encode semantic relationships between words
- d. To generate recommendations
Show Answer
✅ Correct Answer: C
Why it's correct: Embeddings convert words into mathematical vectors so the model can understand how words relate to each other in meaning (e.g., "king" and "queen" are close together).
Why the others are wrong: Embeddings are the "math" behind the language, but they aren't the filters or the specific recommendation engine itself.
- a. Decoder block
- b. Embeddings
- c. Self-attention
- d. Encoder block
Show Answer
✅ Correct Answer: D
Why it's correct: The Encoder block is responsible for processing and understanding the context of the input data.
Why the others are wrong: * Decoder blocks focus on generating the response. * Self-attention is a mechanism inside the blocks, but the Encoder is the overall component.
- a. Image generation
- b. Summarization
- c. Contextual question answering
- d. Personalized recommendations
Show Answer
✅ Correct Answer: C
Why it's correct: Contextual question answering is the specific task of having an AI engage in natural conversation to provide accurate answers based on provided data.
Why the others are wrong: Image generation and Summarization are "one-way" tasks, not conversational dialogues.
- a. Generating new tokens
- b. Detecting anomalies
- c. Analyzing relationships between words from multiple perspectives
- d. Translating between languages
Show Answer
✅ Correct Answer: C
Why it's correct: Multihead attention allows the model to look at different parts of a sentence simultaneously to understand complex relationships (e.g., who "he" refers to in a long paragraph).
Why the others are wrong: It is a specific architectural feature for understanding, not the final act of translating or token generation.
- a. Fast response time
- b. High memory and storage requirements
- c. Low energy consumption
- d. Easy on-premises deployment
Show Answer
✅ Correct Answer: B
Why it's correct: LLMs (like GPT-4) have billions of parameters, requiring massive amounts of GPU memory and storage compared to SLMs (like Phi-3).
Why the others are wrong: SLMs are actually better for fast response times, low energy, and on-premises use.
- a. Fast training times for custom models
- b. Access to a variety of pretrained, high-performance models
- c. Exclusive use of OpenAI models
- d. Only for image-generation tasks
Show Answer
✅ Correct Answer: B
Why it's correct: The Model Catalog gives developers a "one-stop shop" for many different types of state-of-the-art models (OpenAI, Meta, Mistral, etc.).
Why the others are wrong: It isn't just for training or image generation, and it is no longer exclusive to just OpenAI.
- a. It sets user expectations.
- b. It filters out harmful or inappropriate content in real time.
- c. It improves model embeddings.
- d. It provides semantic search capabilities.
Show Answer
✅ Correct Answer: B
Why it's correct: The Safety Layer acts as a filter that checks both the user's prompt and the AI's response for hate speech, violence, or self-harm content.
Why the others are wrong: Setting expectations is a UX task; improving embeddings is a technical training task.
- a. To interpret the context of input
- b. To generate the output sequence based on the encoded input
- c. To embed words into vectors
- d. To attend to specific input tokens
Show Answer
✅ Correct Answer: B
Why it's correct: The Decoder takes the context from the Encoder and predicts the next tokens to create a coherent output.
Why the others are wrong: * Choice A is the Encoder's job. * Choice C is the Embedding layer's job.
- a. Rely solely on automated testing
- b. Avoid documenting potential risks
- c. Use a full-scale rollout immediately
- d. Implement a phased rollout with an incident response plan
Show Answer
✅ Correct Answer: D
Why it's correct: Microsoft recommends a phased rollout (starting with a small group of users) so you can monitor for issues and have a plan ready if the AI behaves unexpectedly.
Why the others are wrong: Full-scale rollouts and ignoring documentation are high-risk behaviors that violate Responsible AI principles.
Goal: Score 80% (44/55) to be exam-ready!
1. A company is developing an AI system to evaluate loan applications for a new fintech startup. Which responsible AI principle should be the primary focus?
- a. Transparency
- b. Fairness
- c. Privacy
- d. Reliability
Show Answer
✅ Correct Answer: B
Why it's correct: Fairness ensures that decisions are made based on relevant financial data and are free from bias or discrimination. In lending, preventing bias against specific groups is the top priority.
Why others are wrong: While transparency (explaining decisions) and privacy (protecting data) are important, they are secondary to the ethical necessity of unbiased lending.
2. A museum wants to make its entire digital archive of historical documents searchable by researchers. Which AI workload is most appropriate?
- a. Computer vision
- b. Knowledge mining
- c. Conversational AI
- d. Sentiment analysis
Show Answer
✅ Correct Answer: B
Why it's correct: Knowledge mining is specifically designed to index large collections of unstructured content (like PDFs and images) and make them searchable.
Why others are wrong: Computer vision sees images, Conversational AI is for chat, and Sentiment analysis identifies emotions—none of these help with deep document indexing.
3. When implementing a new AI system for public use, what is the correct sequence of risk management?
- a. Mitigation, identification, measurement
- b. Identification, mitigation, measurement
- c. Identification, measurement, mitigation
- d. Measurement, identification, mitigation
Show Answer
✅ Correct Answer: C
Why it's correct: You must first identify what could go wrong, then measure how likely/severe it is, and finally mitigate (fix) the risk.
- a. Facial detection
- b. Emotion recognition
- c. Object detection
- d. Language translation
Show Answer
✅ Correct Answer: B
Why it's correct: Microsoft retired Emotion recognition because of concerns regarding its scientific validity and the risk of discriminatory use.
5. A real estate company wants to automatically categorize and analyze property descriptions. Which AI workload best suits this need?
- a. Text analytics
- b. Computer vision
- c. Speech services
- d. Face recognition
Show Answer
✅ Correct Answer: A
Why it's correct: Text analytics (part of Azure AI Language) is designed to extract insights and categories from written property descriptions.
6. In an AI implementation framework, what must be completed before deploying harm-mitigation strategies?
- a. User testing
- b. Impact measurement
- c. System deployment
- d. Marketing analysis
Show Answer
✅ Correct Answer: B
Why it's correct: You cannot effectively fix (mitigate) a problem until you have measured the impact and severity of the potential harm.
7. A retail chain needs to analyze customer feedback across multiple stores. Which AI workload should it use?
- a. Language understanding
- b. Computer vision
- c. Face verification
- d. Speech synthesis
Show Answer
✅ Correct Answer: A
Why it's correct: Language understanding allows the system to process written feedback and interpret the customer's intent or meaning.
8. An autonomous drone delivery service is being developed for medical supplies. Which responsible AI principle should be prioritized?
- a. Inclusiveness
- b. Reliability and safety
- c. Transparency
- d. Fairness
Show Answer
✅ Correct Answer: B
Why it's correct: When dealing with physical drones and medical supplies, ensuring the system operates safely and reliably is the number one priority to prevent accidents.
9. A global hotel chain needs an AI solution to handle customer inquiries in multiple languages 24-7. Which service should it implement?
- a. Conversational AI
- b. Knowledge mining
- c. Image analysis
- d. Text analytics
Show Answer
✅ Correct Answer: A
Why it's correct: Conversational AI (Chatbots) provides real-time, interactive communication needed for 24/7 customer service.
10. Which principle of responsible AI is addressed when an organization documents its AI model’s capabilities and limitations?
- a. Transparency
- b. Accountability
- c. Reliability
- d. Security
Show Answer
✅ Correct Answer: A
Why it's correct: Transparency ensures that people understand how a system works and what its limits are.
11. What is the first step in implementing an AI governance framework for a new health care application?
- a. Deploy security measures
- b. Identify potential harms
- c. Train the model
- d. Test with users
Show Answer
✅ Correct Answer: B
Why it's correct: Identification of potential harms is the foundational first step of the Microsoft Responsible AI standard.
12. A retail analytics company wants to group its shoppers based on browsing patterns to create personalized recommendations. Which type of ML approach is most appropriate?
- a. Supervised learning
- b. Unsupervised learning
- c. Semisupervised learning
- d. Reinforcement learning
Show Answer
✅ Correct Answer: B
Why it's correct: Since the goal is to "group" (cluster) data that doesn't have existing labels, Unsupervised Learning is the correct choice.
13. An agriculture company is building a model to predict crop yields based on soil pH, rainfall, and temperature. What role does temperature play in this model?
- a. Feature
- b. Label
- c. Parameter
- d. Output variable
Show Answer
✅ Correct Answer: A
Why it's correct: Temperature is an input used to make the prediction, which is called a Feature. The yield itself would be the Label.
14. When should you split your dataset into training and validation sets in the Azure Machine Learning workflow?
- a. After model training
- b. During model evaluation
- c. Before model training
- d. During deployment
Show Answer
✅ Correct Answer: C
Why it's correct: You must split the data before training so that the model has "unseen" data (the validation set) to be tested on later.
15. In a health care scenario predicting risk of patient readmission, which metric would be most appropriate for evaluating model performance?
- a. MSE
- b. Accuracy
- c. Silhouette score
- d. R2
Show Answer
✅ Correct Answer: B
Why it's correct: Readmission (Yes/No) is a classification task. Accuracy is a standard metric for classification. (MSE and R2 are for regression).
- a. Poor performance on both training and test data
- b. Equal performance on training and test data
- c. Excellent training performance but poor test performance
- d. Poor training performance but excellent test performance
Show Answer
✅ Correct Answer: C
Why it's correct: Overfitting means the model "memorized" the training data but failed to generalize to the new test data.
- a. Add data transformation modules
- b. Configure compute resources
- c. Create the pipeline infrastructure
- d. Import the dataset
Show Answer
✅ Correct Answer: C
Why it's correct: You have to create the "canvas" or pipeline infrastructure before you can drag and drop modules onto it.
18. A transportation company wants to predict delivery times based on distance, traffic, and weather conditions. Which type of ML problem is this?
- a. Classification
- b. Clustering
- c. Regression
- d. Anomaly detection
Show Answer
✅ Correct Answer: C
Why it's correct: Delivery time is a continuous number, making this a Regression problem.
- a. Training, workspace creation, data preparation, deployment
- b. Data preparation, training, workspace creation, deployment
- c. Workspace creation, data preparation, training, deployment
- d. Deployment, workspace creation, data preparation, training
Show Answer
✅ Correct Answer: C
Why it's correct: You start with the environment (Workspace), then clean data, then train, then deploy.
- a. Overfitting
- b. Underfitting
- c. Data leakage
- d. Perfect fit
Show Answer
✅ Correct Answer: B
Why it's correct: Underfitting happens when the model is too simple to learn the patterns in even the training data.
- a. Before training the model
- b. During model training
- c. After successfully training the model
- d. During data preparation
Show Answer
✅ Correct Answer: C
Why it's correct: An inference pipeline is used to deploy the model and generate predictions. You cannot create it until the model is finalized and "trained."
Why others are wrong: You can't predict (infer) with a model that hasn't finished its training phase yet.
22. An insurance company wants to identify groups of similar claims without predefined categories. Which approach should it use?
- a. Linear regression
- b. Clustering
- c. Binary classification
- d. Time-series analysis
Show Answer
✅ Correct Answer: B
Why it's correct: Clustering is the unsupervised learning technique specifically used to find natural groups in data without labels.
23. A global hotel chain needs to detect guest comments in various languages and quickly route them to appropriate departments. Which Azure AI Service should it use first?
- a. Text analytics
- b. Language detection
- c. Sentiment analysis
- d. Entity recognition
Show Answer
✅ Correct Answer: B
Why it's correct: Before you can analyze or translate a comment, you must first detect which language it is written in to route it correctly.
24. A pharmaceutical company wants to extract mentions of medical conditions, treatments, and dosages from clinical trial documents. Which service is most appropriate?
- a. Key phrase extraction
- b. Text analytics
- c. Named entity recognition (NER)
- d. Sentiment analysis
Show Answer
✅ Correct Answer: C
Why it's correct: NER identifies and categorizes specific "entities" like diseases or medications. Azure AI Language also has a specialized version called "NER for Health."
25. What would be the output of Azure AI Language Detection when processing text in an unsupported language?
- a. Empty string
- b. Error code
- c. NaN confidence score
- d. Default to English
Show Answer
✅ Correct Answer: C
Why it's correct: If the language isn't recognized, the system returns NaN (Not a Number) for the confidence score because it cannot calculate how certain it is.
26. An international law firm needs to translate legal contracts while maintaining their formatting and structure. Which translation approach should it use?
- a. Real-time translation
- b. Asynchronous batch translation
- c. Custom neural voice
- d. Text-to-speech translation
Show Answer
✅ Correct Answer: B
Why it's correct: Batch translation (Document Translation) is designed to translate entire files (like Word or PDF) while keeping the original layout, tables, and fonts.
27. A startup is creating a virtual assistant that needs to understand user requests in multiple languages. Which service combination should it use?
- a. Language understanding and Translator
- b. Speech Service and Face API
- c. Text analytics and computer vision
- d. Form recognizer and Translator
Show Answer
✅ Correct Answer: A
Why it's correct: A virtual assistant needs Language understanding to know what the user wants (intent) and Translator to handle various languages.
- a. Entity linking
- b. Language detection
- c. Content filtering
- d. Speech recognition
Show Answer
✅ Correct Answer: C
Why it's correct: Content filtering uses AI to block hate speech, violence, and self-harm content in real-time.
29. An educational platform needs to create audiobooks in multiple languages with natural-sounding voices. Which service should it use?
- a. Text analytics
- b. Language understanding
- c. Custom neural voice
- d. Entity recognition
Show Answer
✅ Correct Answer: C
Why it's correct: Custom neural voice allows you to create highly realistic, human-like speech from text.
30. Which component is essential for creating a custom translation model for industry-specific terminology?
- a. Large general dataset
- b. Parallel sentence pairs
- c. Speech samples
- d. Image annotations
Show Answer
✅ Correct Answer: B
Why it's correct: To train a custom translator, you need parallel sentence pairs (the same sentence in two different languages) so the model learns the specific mapping.
31. A media company needs to automatically generate subtitles for live broadcasts. Which service should it use?
- a. Text analytics
- b. Speech-to-text
- c. Custom Translator
- d. Language understanding
Show Answer
✅ Correct Answer: B
Why it's correct: Speech-to-text transcribes spoken audio into written words, which is exactly how subtitles are made.
32. Which Azure AI service would help a company compare customer support tickets for similarity to avoid duplicate efforts?
- a. Sentiment analysis
- b. Language detection
- c. Text embeddings
- d. NER
Show Answer
✅ Correct Answer: C
Why it's correct: Embeddings represent text as math vectors; if two vectors are close together, the tickets are semantically similar.
33. A telehealth company needs to identify different speakers in recorded medical consultations. Which service should it use?
- a. Language detection
- b. Text analytics
- c. Speaker recognition
- d. Entity linking
Show Answer
✅ Correct Answer: C
Why it's correct: Speaker recognition (Diarization) identifies who is speaking in an audio file.
34. A wildlife conservation organization needs to count different species of animals in drone footage. Which vision service is most appropriate?
- a. Image classification
- b. Object detection
- c. OCR
- d. Facial detection
Show Answer
✅ Correct Answer: B
Why it's correct: Object detection identifies what the animal is and where it is, allowing you to count multiple animals in one frame.
35. A smart parking system needs to identify available spaces by analyzing whether each parking spot is empty or occupied. Which approach is most suitable?
- a. Face recognition
- b. OCR
- c. Semantic segmentation
- d. Landmark detection
Show Answer
✅ Correct Answer: C
Why it's correct: Semantic segmentation labels every pixel in an image (e.g., "Car" or "Empty Asphalt"), which is perfect for precise boundary detection in a parking lot.
- a. Image metadata only
- b. Object size only
- c. Bounding box coordinates and confidence score
- d. Color information only
Show Answer
✅ Correct Answer: C
Why it's correct: You get the Bounding box (location) and the Confidence score (how sure the AI is).
37. A museum wants to identify famous paintings and provide information about them. Which specialized domain model should it use?
- a. Celebrity recognition
- b. Custom vision
- c. Facial detection
- d. OCR
Show Answer
✅ Correct Answer: B
Why it's correct: Since "Paintings" isn't a standard pre-built category like celebrities, the museum should use Custom Vision to train its own model.
- a. Facial detection
- b. Facial verification
- c. Emotion recognition
- d. Face location
Show Answer
✅ Correct Answer: C
Why it's correct: Microsoft retired Emotion recognition because it is ethically sensitive and often scientifically unreliable.
39. A financial institution needs to process handwritten loan applications. Which service combination should it use?
- a. Form recognizer with OCR
- b. Facial detection with OCR
- c. DALL-E with OCR
- d. Landmark detection with OCR
Show Answer
✅ Correct Answer: A
Why it's correct: Form Recognizer (Document Intelligence) is the best tool for extracting data from structured forms, even if they are handwritten.
- a. Face recognition
- b. Image variation generation
- c. Text extraction
- d. Object counting
Show Answer
✅ Correct Answer: B
Why it's correct: You can give DALL-E an image and ask it to create variations of that image (different style, lighting, etc.).
41. A security system needs to verify if a person is physically present versus using a photo. Which feature should be used?
- a. Facial detection
- b. Liveness detection
- c. Object detection
- d. Image classification
Show Answer
✅ Correct Answer: B
Why it's correct: Liveness detection is a specific feature of Azure AI Face that distinguishes between a "live" person and a spoof (like a printed photo or a mask).
42. A historical archive wants to convert old handwritten letters into searchable text. Which capability is most appropriate?
- a. Image classification
- b. Object detection
- c. Handwritten text recognition
- d. Facial detection
Show Answer
✅ Correct Answer: C
Why it's correct: Azure AI Vision's **Read API** includes specialized Handwritten text recognition to digitize non-printed text.
- a. Existing image only
- b. Natural language description
- c. Programming code
- d. Audio file
Show Answer
✅ Correct Answer: B
Why it's correct: DALL-E is a "text-to-image" model; it generates visuals based on a natural language prompt (e.g., "A cat wearing a tuxedo").
44. A retail store needs to analyze customers’ movement patterns without identifying individuals. Which vision service is appropriate?
- a. Face recognition
- b. OCR
- c. Spatial analysis
- d. Landmark detection
Show Answer
✅ Correct Answer: C
Why it's correct: Spatial analysis tracks how people move through a physical space (heatmaps, line counting) without needing to identify who they are.
45. A software company needs to generate test cases for its REST APIs. Which generative AI model would be most appropriate?
- a. DALL-E
- b. GPT
- c. Whisper
- d. Stable Diffusion
Show Answer
✅ Correct Answer: B
Why it's correct: GPT (Generative Pre-trained Transformer) models are designed for text and code generation, making them perfect for creating test cases.
46. A marketing team wants to integrate AI assistance into email campaign writing. Which solution best fits their needs?
- a. Chatbot
- b. Copilot
- c. Translation service
- d. Content Moderator
Show Answer
✅ Correct Answer: B
Why it's correct: Copilot solutions are designed to assist with content creation in specific domains like marketing, helping to draft text alongside a human user.
- a. API key
- b. Tone specification
- c. Model version
- d. Response length
Show Answer
✅ Correct Answer: B
Why it's correct: A Tone specification (e.g., "You are a formal assistant") in the System Message guides the AI's stylistic output.
- a. User interface layer
- b. Model layer
- c. Safety system layer
- d. Network layer
Show Answer
✅ Correct Answer: C
Why it's correct: In Azure OpenAI's architecture, the Safety system layer is where real-time filtering for harmful content takes place.
49. A development team needs to auto-generate database queries. Which programming language does the AI model best support?
- a. Assembly
- b. Python
- c. COBOL
- d. Machine code
Show Answer
✅ Correct Answer: B
Why it's correct: Python is one of the most widely supported languages in modern LLMs due to its massive presence in training datasets.
50. What approach should be used when the AI model needs to generate content without any previous examples?
- a. Zero-shot learning
- b. Supervised learning
- c. Transfer learning
- d. Reinforcement learning
Show Answer
✅ Correct Answer: A
Why it's correct: Zero-shot learning is when a model performs a task based only on its pre-training, without being given any specific examples in the prompt.
51. Which component is essential for implementing safe deployment of a generative AI system in health care?
- a. Real-time translation
- b. Content safety filters
- c. Image generation
- d. Speech synthesis
Show Answer
✅ Correct Answer: B
Why it's correct: Content safety filters are critical in sensitive fields like healthcare to ensure the AI doesn't provide harmful or inappropriate medical advice.
- a. Complete system replacement
- b. Repetitive task automation
- c. Security auditing
- d. Performance optimization
Show Answer
✅ Correct Answer: B
Why it's correct: Copilots are "assistants." They excel at repetitive tasks (like writing boilerplate code), allowing the human to focus on complex logic.
- a. Changing API keys
- b. Using system messages
- c. Increasing model size
- d. Modifying network settings
Show Answer
✅ Correct Answer: B
Why it's correct: System messages set the "ground rules" and behavior instructions that stay active across the entire conversation.
- a. Web interface only
- b. Domain-specific training
- c. Public dataset
- d. Social media integration
Show Answer
✅ Correct Answer: B
Why it's correct: For niche industries, the model needs domain-specific context (often via RAG or Fine-tuning) to be accurate and useful.
- a. Speed optimization
- b. Grounding with reference data
- c. Network configuration
- d. User interface design
Show Answer
✅ Correct Answer: B
Why it's correct: Grounding (using RAG) forces the AI to look at "reference data" (your own trusted documents) before answering, which significantly reduces made-up facts (hallucinations).