AI-900 Hard Questions — 25 Difficult Questions

Question 1

A medical research team wants to analyze MRI scans to identify and measure the precise boundaries of tumors. They need to assign each pixel in the image to a class (e.g., tumor, healthy tissue, background). Which Azure Computer Vision capability should they use?

Accepted Answer

Semantic Segmentation. Semantic segmentation assigns a class label to every pixel in an image, making it the correct choice for precisely delineating tumor boundaries in MRI scans. Azure Computer Vision's semantic segmentation capability outputs a pixel-level mask, enabling the research team to differentiate tumor, healthy tissue, and background at the finest granularity.

Answer

Object Detection

Answer

Image Classification

Answer

Optical Character Recognition

Question 2

A quality assurance team at a software company uses Azure OpenAI Service to generate compliance reports. They need the model to produce the exact same output for a given prompt every time the API is called, to ensure reproducibility during testing. Which parameter should they set to achieve this deterministic behavior?

Accepted Answer

Set temperature to 0. Setting temperature to 0 forces the model to choose the most likely token at each step, eliminating randomness and producing deterministic outputs for the same prompt. This is essential for reproducibility in testing scenarios where identical results are required across API calls.

Answer

Set frequency penalty to 1

Answer

Set top_p to 1

Answer

Set max_tokens to the expected output length

Question 3

A data scientist trains a regression model to predict house prices using features like bedrooms, square footage, and location. The model achieves an R-squared of 0.95 on the test set. However, when deployed to predict prices in a new city with different property characteristics, the predictions are very inaccurate. Which concept best explains this poor performance?

Accepted Answer

Overfitting. The model achieved an R-squared of 0.95 on the test set, indicating excellent performance on data from the same distribution. However, when deployed to a new city with different property characteristics, the predictions were very inaccurate. This is a classic symptom of overfitting, where the model has learned noise and patterns specific to the training data (e.g., city-specific price trends) that do not generalize to unseen data from a different distribution.

Answer

Underfitting

Answer

High bias

Answer

Data drift

Question 4

A manufacturing company wants to use Azure Computer Vision to inspect products on an assembly line for defects. They have a labeled dataset with images of defective and non-defective products. They need to not only classify products as defective or not, but also identify the exact location of the defect (e.g., a crack) in the image. Which Azure Computer Vision capability should they use?

Accepted Answer

Custom Vision object detection. Custom Vision object detection is the correct choice because it not only classifies images (defective vs. non-defective) but also localizes defects by drawing bounding boxes around them. The labeled dataset with defect locations directly supports training a model to output both class labels and spatial coordinates, which is exactly what object detection provides.

Answer

Custom Vision image classification

Answer

Azure Face API

Answer

Optical Character Recognition (OCR)

Question 5

A medical research team needs to analyze CT scans to identify and outline the exact boundaries of lung nodules. Which Azure Computer Vision capability should they use?

Accepted Answer

Semantic Segmentation. Semantic segmentation is the correct capability because it classifies each pixel in an image, enabling precise delineation of object boundaries. For CT scans, this allows the model to outline the exact shape and contour of lung nodules, which is essential for medical analysis. Image classification and object detection only provide labels or bounding boxes, not pixel-level boundaries.

Answer

Image Classification

Answer

Object Detection

Answer

Optical Character Recognition (OCR)

Question 6

A medical research organization needs to process thousands of clinical trial documents to automatically extract specific medical entities such as diseases, symptoms, medications, and dosages. They want to use a prebuilt Azure AI Language capability that is already trained on medical domain data. Which Azure AI Language feature should they use?

Accepted Answer

Text Analytics for Health. Text Analytics for Health is a prebuilt Azure AI Language capability specifically trained on medical domain data, enabling extraction of entities like diseases, symptoms, medications, and dosages from clinical trial documents without requiring custom model training. It is designed for healthcare and life sciences use cases, making it the correct choice for processing thousands of clinical documents automatically.

Answer

Key Phrase Extraction

Answer

Named Entity Recognition (NER)

Answer

Custom Text Classification

Question 7

An autonomous vehicle team needs a system that not only identifies objects like cars and pedestrians but also creates a precise pixel-level mask for each individual object instance, even when objects overlap. Which Azure Computer Vision capability should they use?

Accepted Answer

Instance segmentation. Instance segmentation (Option D) is the correct choice because it combines object detection with semantic segmentation to identify each individual object instance and generate a precise pixel-level mask for it, even when objects overlap. This capability is essential for autonomous vehicles to distinguish between multiple cars or pedestrians that may partially occlude each other, enabling safe navigation.

Answer

Image classification

Answer

Object detection

Answer

Semantic segmentation

Question 8

What is the 'bias-variance tradeoff' in machine learning?

Accepted Answer

The tradeoff between underfitting (high bias) and overfitting (high variance) when choosing model complexity. Option B is correct because the bias-variance tradeoff describes the inverse relationship between underfitting (high bias, where the model is too simple to capture patterns) and overfitting (high variance, where the model is too complex and captures noise). In Azure Machine Learning, this tradeoff is managed by tuning hyperparameters like regularization strength or tree depth to balance model complexity and generalization.

Answer

The tradeoff between model accuracy and inference speed

Answer

The tradeoff between training data quantity and model quality

Answer

The difference in fairness metrics between biased and unbiased model versions

Question 9

A bank deploys an AI system that uses a deep neural network to approve personal loan applications. A customer whose loan was rejected requests a detailed explanation of why the decision was made. The bank's AI team realizes that the model's internal workings are too complex to provide a simple, understandable reason. According to Microsoft's responsible AI principles, which principle is most directly violated by this situation?

Accepted Answer

Transparency. The bank's inability to provide a clear, understandable explanation for the AI's loan decision directly violates the transparency principle. Microsoft's responsible AI principles require that AI systems be understandable and that their decisions can be explained to users, especially when those decisions have significant impact. A deep neural network's complex, non-linear decision boundaries and lack of inherent interpretability make it a 'black box,' which undermines the required transparency.

Answer

Fairness

Answer

Reliability & Safety

Answer

Privacy & Security

Question 10

A customer support team receives thousands of emails daily. They want to automatically route each email to the appropriate department (Billing, Technical Support, or General Inquiry). They also want to extract the customer's account number and order ID from each email. Which combination of Azure AI Language features should they use?

Accepted Answer

Text classification and custom entity recognition. Option C is correct because the scenario requires two distinct NLP tasks: categorizing emails into predefined departments (Billing, Technical Support, General Inquiry) which is a text classification task, and extracting specific structured data (account number and order ID) which requires custom entity recognition to identify domain-specific entities not covered by prebuilt NER. Azure AI Language provides both custom text classification and custom entity recognition features to handle these requirements.

Answer

Sentiment analysis and key phrase extraction

Answer

Language detection and translation

Answer

Named entity recognition (NER) and summarization

Question 11

A data science team trains a regression model to predict house prices. They evaluate the model using Mean Absolute Error (MAE). After deployment, they notice that the model occasionally produces large errors (e.g., underpredicting a luxury home by $500,000) while most predictions are within $20,000. The business is more concerned about the impact of these large errors than the average small error. Which additional metric should the team use to better capture the penalty for large errors?

Accepted Answer

Root Mean Squared Error (RMSE). Root Mean Squared Error (RMSE) is the correct additional metric because it squares the residuals before averaging, which heavily penalizes large errors like the $500,000 underprediction. Unlike MAE, which treats all errors equally, RMSE amplifies the impact of outliers, making it a better fit for a business that cares more about catastrophic failures than typical small errors. This aligns with the need to capture the penalty for large deviations in regression model evaluation.

Answer

R-squared

Answer

Mean Absolute Percentage Error (MAPE)

Answer

F1 score

Question 12

A city deploys an AI-powered kiosk to help residents access government services. The kiosk uses a voice interface only, without any text or screen reader support. Which Microsoft responsible AI principle is most directly being ignored?

Accepted Answer

C. The kiosk uses only a voice interface without text or screen reader support, which directly violates the Microsoft responsible AI principle of Inclusiveness. Inclusiveness requires that AI systems are designed to empower everyone, including people with disabilities such as hearing impairments or those who rely on visual or text-based interfaces. By excluding non-verbal interaction methods, the system fails to accommodate diverse user needs, making it inaccessible.

Answer

A

Answer

B

Answer

D

Question 13

A company deploys an AI-powered voice assistant that only supports English. The assistant is used in a country where the official languages are English, French, and Dutch. Many users who speak French or Dutch cannot use the assistant effectively. Which Microsoft responsible AI principle is most directly relevant to this situation?

Accepted Answer

Inclusiveness. The assistant's inability to support French and Dutch users directly violates the inclusiveness principle, which requires AI systems to be designed for all users regardless of language, ability, or background. By supporting only English in a multilingual country, the system excludes a significant portion of the target audience, failing to provide equitable access.

Answer

Fairness

Answer

Reliability and safety

Answer

Transparency

Question 14

A company develops an AI system to screen job candidates based on their resumes. The system is trained on historical data. Analysis reveals that the model has an adverse impact against female candidates due to a proxy feature (e.g., 'years of continuous employment') that correlates with gender. The team removes the protected attribute 'gender' from the training data but the biased outcome persists. According to Microsoft's responsible AI principles, which additional step should the team take to address this unfairness?

Accepted Answer

Use a tool like Fairlearn to detect and mitigate the bias while maintaining model performance.. Option B is correct because Microsoft's responsible AI principle of fairness requires not just removing protected attributes but also detecting and mitigating proxy features that cause bias. Fairlearn is a Microsoft open-source toolkit specifically designed to assess and mitigate unfairness in AI systems, offering algorithms like 'Exponentiated Gradient Reduction' or 'Grid Search' to reduce disparity while preserving model performance. Simply removing the proxy feature (A) may not always be feasible if it carries predictive value, and Fairlearn provides a systematic way to balance fairness and accuracy.

Answer

Remove the offending proxy feature 'years of continuous employment' from the training data.

Answer

Train a separate model for each gender group to ensure equal outcomes.

Answer

Collect more training data from underrepresented groups.

Question 15

A data scientist has a small dataset with only 200 labeled samples. They want to get a reliable estimate of model performance without using a separate validation set that would reduce the training data. Which technique should the data scientist use in Azure Machine Learning to obtain this reliable estimate?

Accepted Answer

k-fold cross-validation. B is correct because k-fold cross-validation splits the small dataset into k folds, trains the model on k-1 folds, and validates on the remaining fold, repeating this process k times. This provides a reliable performance estimate by using all 200 samples for both training and validation without requiring a separate hold-out set, which is critical for small datasets in Azure Machine Learning.

Answer

Hold-out validation

Answer

Data augmentation

Answer

Principal Component Analysis (PCA)

Question 16

A company develops an AI system to screen job resumes and rank candidates for interviews. The system is trained on historical hiring data that favored candidates from certain well-known universities. The company decides to deploy the system without any adjustments to address this bias. Which Microsoft responsible AI principle is most directly being violated?

Accepted Answer

Fairness. The correct answer is A (Fairness) because the AI system was trained on biased historical data that favored candidates from certain universities, and deploying it without adjustments directly violates the fairness principle. Fairness in responsible AI requires that systems treat all people equitably and do not discriminate based on protected attributes such as educational background. By not mitigating the bias, the system perpetuates historical inequities in the hiring process.

Answer

Inclusiveness

Answer

Reliability and Safety

Answer

Privacy and Security

Question 17

A company uses a GPT-based model to generate marketing copy. They notice the model occasionally produces text that includes harmful stereotypes. They want to reduce these harmful outputs without retraining the model. Which approach is most appropriate?

Accepted Answer

Prompt engineering with specific instructions to avoid stereotypes. Option B is correct because prompt engineering allows you to guide the model's behavior at inference time without modifying its weights. By including explicit instructions in the prompt (e.g., 'Avoid harmful stereotypes'), you can steer the output toward safer content. This is the most appropriate approach when retraining is not an option, as it directly addresses the undesired outputs through input design.

Answer

Fine-tuning the model on a curated dataset

Answer

Reducing the temperature parameter to zero

Answer

Increasing the maximum output length

Question 18

A company uses a generative AI model to create blog posts. They want to ensure that the model's output never contains offensive or harmful language before the content is published. They implement a system that checks the generated text against a list of prohibited terms and blocks or edits the content if necessary. Which type of safety measure is this?

Accepted Answer

Post-processing content filtering. Option C is correct because the described system operates after the model generates text, scanning the output against a prohibited terms list and blocking or editing it. This is a classic post-processing content filtering approach, distinct from modifying the model's training data, prompts, or weights. Azure AI Content Safety is an example of such a post-processing filter that can be applied to generative AI outputs.

Answer

Pre-training data cleaning

Answer

Prompt engineering with safety instructions

Answer

Model fine-tuning on safe examples

Question 19

A data scientist is training a binary classification model to detect fraudulent transactions. The dataset contains only 1% fraudulent transactions. The model achieves 99% accuracy on the test set, but when deployed, it fails to detect most actual fraud cases. Which metric would best reveal this issue?

Accepted Answer

Recall. Recall (sensitivity) measures the proportion of actual positive cases correctly identified. In this highly imbalanced dataset (1% fraud), a model can achieve 99% accuracy by simply predicting 'non-fraud' for every transaction, which yields zero true positives. Recall reveals this failure because it focuses solely on how many fraudulent transactions were caught, ignoring the vast majority of non-fraud cases.

Answer

Accuracy

Answer

Precision

Answer

F1 score

Question 20

A data scientist is building a classification model to detect fraudulent transactions. The dataset has 1,000,000 legitimate transactions and only 1,000 fraudulent ones. The model achieves 99.9% accuracy on the test set, but it fails to catch most fraudulent cases. Which metric should the data scientist prioritize to better evaluate the model's performance on this imbalanced dataset?

Accepted Answer

Recall. Recall measures the proportion of actual positive cases (fraudulent transactions) correctly identified by the model. With only 1,000 fraud cases out of 1,001,000 total transactions, a model that predicts 'legitimate' for every transaction would achieve 99.9% accuracy but 0% recall, making recall the critical metric for imbalanced fraud detection.

Answer

Accuracy

Answer

Mean Squared Error

Answer

R-squared