Practice for the AI-900: Microsoft Azure AI Fundamentals exam with 15 exam-style practice questions, instant answer reveals, and concise explanations of every correct answer. Topics include: Why is it important to divide the available data into training, validation, and test sets?. Follow @CertPunch and visit certpunch.com for more certification practice exams and study content.
Prefer hands-on? Take this round as an interactive practice test — answer every question, get instant feedback, and see your score: Start the AI-900: Microsoft Azure AI Fundamentals practice test →
What you will practice
- Why is it important to divide the available data into training, validation, and test sets?
- An organization has deployed a deep learning model using Azure AI services. Soon after deployment, they obser…
- Which statement best describes the primary functionality of Optical Character Recognition (OCR) in Azure AI V…
- An organization is setting up a language service resource to leverage multiple NLP capabilities (e.g., transl…
- Which feature is designed to enhance security by distinguishing between a live face and a counterfeit represe…
- A company wants to build a hotel booking chatbot that identifies user goals (e.g., "book a room," "cancel res…
Answers and explanations
Tap a question to expand the answer and the exam reasoning. Try to commit to your own pick first.
Q1. Why is it important to divide the available data into training, validation, and test sets?
Answer: D. To train the model on one set, adjust hyperparameters with a separate set, and finally assess unbiased performance on unseen data.
Splitting data lets the model learn from the training set, tune hyperparameters using the validation set, and provide an unbiased evaluation on the test set. Using the training data for final evaluation would yield overly optimistic metrics because the model has already seen those examples.
Q2. An organization has deployed a deep learning model using Azure AI services. Soon after deployment, they observe that the model's outcomes are consistently skewed toward a particular demographic. Which of the following approaches is least e…
Answer: B. Rely solely on the model's internal decision‑making transparency to correct bias.
Deep learning models are largely opaque, so relying on their internal transparency yields little actionable insight for correcting demographic skew. The strongest distractors, such as fairness audits and dataset augmentation, are actually effective mitigation steps, making the transparency option the clear outlier.
Q3. Which statement best describes the primary functionality of Optical Character Recognition (OCR) in Azure AI Vision?
Answer: B. It converts images into machine-readable text by leveraging both visual analysis and text interpretation.
Optical Character Recognition extracts text from images and converts it into a machine-readable digital format for further processing. It is not a tool for enhancing image visualization; instead, expect questions to frame OCR as pulling printed or handwritten text out of pictures.
Q4. An organization is setting up a language service resource to leverage multiple NLP capabilities (e.g., translation, text analytics, and text-to-speech) within a single Azure AI service resource. What is a key benefit of using a multi-servi…
Answer: C. It consolidates billing and management for multiple language services under one resource.
A multi-service resource consolidates billing and management for various Azure AI services under a single endpoint and subscription key. This approach simplifies administration compared to provisioning individual resources, which would require tracking separate endpoints and keys.
Q5. Which feature is designed to enhance security by distinguishing between a live face and a counterfeit representation, such as a printed photograph or mask?
Answer: D. Face liveness detection
Face liveness detection evaluates physical presence to prevent spoofing using photographs or masks. Face recognition simply matches identities, so it fails to verify if the source is an actual live human.
Q6. A company wants to build a hotel booking chatbot that identifies user goals (e.g., "book a room," "cancel reservation") and extracts details like dates and locations. Which Azure AI Language component is most appropriate for this task? Cor…
Answer: A. Conversational Language Understanding (CLU)
Conversational Language Understanding predicts user intents and extracts entities like dates from dialogue, making it the standard tool for chatbots. Key phrase extraction only highlights main topics without understanding the underlying goal.
Q7. What distinguishes face verification from face identification in Azure Face services?
Answer: A. Face verification performs a one-to-one comparison, whereas face identification compares a face against a group to find a match.
Face verification performs a one-to-one comparison to confirm identity, while face identification performs a one-to-many comparison against a database. Remember verification checks if two faces match, whereas identification searches a group.
Q8. A medical team is building a cancer detection model where missing positive cases could be life-threatening. Which metric should they prioritize?
Answer: D. Recall
Recall measures the proportion of actual positive cases correctly identified, which is critical for minimizing false negatives in life-threatening medical scenarios. Maximizing recall ensures the model captures all potential illnesses, even if false positives increase.
Q9. Which statement best describes overfitting in a machine learning model?
Answer: B. The model captures noise and irrelevant details from the training data, resulting in high performance on training data but poor performance on new data.
Overfitting happens when a model learns noise and irrelevant details from the training data too closely, failing to generalize. This results in excellent training performance but poor accuracy on new data, limiting its real-world usefulness.
Q10. A retail company wants to quickly analyze customer reviews to determine whether the sentiment expressed is positive or negative without investing time in model training. Which Azure AI language capability should they use?
Answer: C. Prebuilt sentiment analysis
Prebuilt sentiment analysis provides immediate results without requiring custom model training, making it ideal for quick text evaluation. Custom solutions require training, whereas prebuilt models instantly categorize text as positive, negative, or neutral.
Q11. A company is implementing an AI solution that needs to understand the relationships between medical terms in patient records. Which characteristic of large language models would be most beneficial for this requirement?
Answer: B. Vector embeddings
Vector embeddings map words into a mathematical space where semantically similar terms cluster together, which captures relationships between concepts. For your exam, remember that prompt engineering shapes the input, while embeddings represent the actual meaning of the text.
Q12. An insurance company needs to process 500 legacy claim forms in various layouts while preserving original tables and formatting. Which Azure AI Translator configuration is most appropriate?
Answer: A. Asynchronous batch translation with blob storage
Asynchronous batch translation handles multiple documents simultaneously while preserving their original structures and formatting. While the REST API offers document translation, batch processing with storage is the most robust choice for heavy enterprise workloads.
Q13. A healthcare organization has developed an AI model to diagnose diseases from medical images. Despite high accuracy on training images, the model fails to generalize well on new data due to overfitting. Which of the following strategies is…
Answer: C. Adding additional layers to further increase model complexity
Increasing model complexity will only exacerbate overfitting because the model memorizes the noise in the training data instead of generalizing. As a practical study cue, focus on techniques like adding data and early stopping to improve generalization.
Q14. Which risk associated with AI does conducting regular fairness audits on training data primarily address?
Answer: C. The risk of producing biased outcomes due to unrepresentative or skewed data.
Fairness audits directly address the risk of biased outcomes by identifying unrepresentative or skewed data before model training. Do not confuse this with security, which handles sensitive data storage, or opacity, which deals with how deep networks make decisions.
More AI-900: Microsoft Azure AI Fundamentals drills and other practice exams are on @CertPunch. New rounds drop every few days at certpunch.com.