AI-900 Hard Questions — 25 Difficult Questions

Question 1

What is 'image generation quality' evaluation — how do you measure if a generated image is good?

Accepted Answer

Metrics like FID (image distribution similarity) and CLIP score (prompt adherence), plus human evaluation. Image generation quality is evaluated using a combination of automated metrics and human judgment. FID (Fréchet Inception Distance) measures how similar the distribution of generated images is to real images, while CLIP score assesses how well the image aligns with the given text prompt. Human evaluation is also critical to capture perceptual quality that automated metrics may miss, such as aesthetic appeal and contextual coherence.

Answer

Only image resolution and file size — higher resolution means better quality

Answer

Simply asking the model what score it gives its own output

Answer

Counting the number of objects correctly included vs. missing from the prompt

Question 2

What is 'coreference resolution' in natural language processing?

Accepted Answer

Identifying which words or phrases in a text refer to the same real-world entity. Coreference resolution is the NLP task of identifying when two or more expressions in a text refer to the same real-world entity. For example, in 'Alice said she would come,' the pronoun 'she' corefers to 'Alice.' This is fundamental for tasks like document summarization and question answering, where maintaining entity consistency is critical.

Answer

Checking whether a document's references (citations) are correctly formatted

Answer

Resolving conflicts when multiple languages are mixed in the same document

Answer

Matching database foreign keys to their referenced primary keys

Question 3

A data scientist trains a regression model to predict energy consumption for a smart building. The model achieves very low error on the training data but performs significantly worse on a held-out validation set. Which technique would most directly address this problem?

Accepted Answer

Regularization. The model's low training error but high validation error indicates overfitting, where the model has memorized the training data rather than learning generalizable patterns. Regularization (e.g., L1 or L2) directly penalizes large coefficients, reducing model complexity and improving generalization to unseen data.

Answer

Feature engineering

Answer

Cross-validation

Answer

Hyperparameter tuning

Question 4

A retail company wants to use security cameras to analyze customer flow. They need to detect when a person enters a specific store zone, count how many people are in that zone at any given time, and track the direction each person moves within the zone. Which Azure Computer Vision capability should they use?

Accepted Answer

Spatial Analysis. Spatial Analysis is the correct Azure Computer Vision capability because it is specifically designed to analyze video feeds from cameras to detect people, count them in defined zones, and track their movement direction. Unlike general object detection, Spatial Analysis provides the specialized functions for zone occupancy and person trajectory tracking required by the retail scenario.

Answer

Object detection

Answer

Optical Character Recognition (OCR)

Answer

Semantic segmentation

Question 5

What is the 'dual-use' problem in AI and why is it relevant to responsible deployment?

Accepted Answer

The risk that AI capabilities designed for good can also be used for harmful purposes. The 'dual-use' problem in AI refers to the risk that a technology designed for beneficial purposes can also be misapplied for harmful ends. This is central to responsible deployment because it forces organizations to consider not only the intended use case but also potential misuse, such as facial recognition systems used for surveillance or generative AI creating disinformation. Addressing dual-use requires implementing safeguards like usage policies, access controls, and ethical review boards.

Answer

When an AI model is licensed for use by two different organisations simultaneously

Answer

Deploying the same AI model for both training and inference to reduce costs

Answer

Combining two AI models to achieve better results than either model alone

Question 6

A law firm needs to automatically extract specific information from legal contracts, such as the names of the parties involved, effective dates, and governing law clauses. The firm has a small set of contracts that have been manually annotated with these specific fields. Which Azure AI Language feature should they use to build a custom extraction solution?

Accepted Answer

Custom Named Entity Recognition (NER). Custom Named Entity Recognition (NER) is the correct choice because the law firm needs to extract specific, custom fields (party names, effective dates, governing law clauses) from legal contracts, which are not covered by prebuilt entity categories. Custom NER allows you to train a model using a small set of manually annotated contracts to recognize these domain-specific entities, enabling tailored extraction for the firm's unique requirements.

Answer

Prebuilt Named Entity Recognition (NER)

Answer

Key Phrase Extraction

Answer

Text Summarization

Question 7

A parking lot management company uses security cameras to monitor vehicles. They need to both detect the presence of license plates in an image and read the alphanumeric characters on those plates. Which Azure Computer Vision capability should they use to achieve both requirements?

Accepted Answer

Optical Character Recognition (OCR) - Read API. (OCR - Read API) is correct because Azure's Read API is specifically designed to both detect the presence of text (including license plates) in an image and extract the alphanumeric characters from that text. This meets both requirements—detecting the plate and reading its characters—in a single call, using deep-learning-based recognition models optimized for printed and handwritten text.

Answer

Image Analysis (describe image and detect objects)

Answer

Face API

Answer

Custom Vision (object detection)

Question 8

A financial company develops an AI system that recommends loan amounts based on historical data. The historical data includes years of discriminatory lending practices against certain minority groups. As a result, the AI system disproportionately denies loans to members of those groups. Which Microsoft responsible AI principle is most directly violated by this scenario?

Accepted Answer

Fairness. The AI system's loan recommendations are based on historical data that contains discriminatory lending practices, leading to disproportionate denials for minority groups. This directly violates the Fairness principle, which requires AI systems to treat all people equitably and avoid reinforcing existing biases. The system's outputs are not fair because they perpetuate historical inequities, making fairness the most relevant principle.

Answer

Reliability and Safety

Answer

Privacy and Security

Answer

Inclusiveness

Question 9

A company deploys an AI system to screen job resumes. The system consistently rejects candidates from a certain university, but the company cannot determine which features led to the decision or how the model arrived at that outcome. Which Microsoft responsible AI principle is most directly violated?

Accepted Answer

Transparency. The scenario describes a system that makes decisions without providing any insight into how or why those decisions were made. Transparency, as a Microsoft responsible AI principle, requires that AI systems be understandable and that their decisions can be explained. Since the company cannot determine which features led to the rejection or how the model arrived at the outcome, the lack of explainability directly violates the Transparency principle.

Answer

Fairness

Answer

Accountability

Answer

Privacy and security

Question 10

What is 'model cards' in responsible AI and what information do they contain?

Accepted Answer

Transparency documents describing a model's intended use, training data, performance, biases, and limitations. Model cards are transparency documents that accompany machine learning models to disclose their intended use, training data, performance metrics, known biases, and limitations. They are a key responsible AI practice, mandated by frameworks like Microsoft's Responsible AI Standard, to ensure stakeholders understand a model's capabilities and risks before deployment.

Answer

Azure billing documents showing the monthly cost of running a model in production

Answer

Technical specification sheets for AI hardware accelerators used in model training

Answer

Playing cards used in gamification of AI training to motivate data labellers

Question 11

A medical research organization uses an AI system to analyze patient health records to identify patterns in disease progression. They publish a research paper that includes tables of aggregated statistics derived from the data. Later, a researcher discovers that by combining multiple statistics, it is possible to identify individual patients. Which Microsoft responsible AI principle has been most directly compromised?

Accepted Answer

Privacy and security. The scenario describes a re-identification attack, where aggregated statistics (tables) can be combined to infer individual patient identities. This directly violates the privacy and security principle, which requires that AI systems protect personal data and prevent unauthorized identification. Microsoft's responsible AI principle of privacy and security emphasizes safeguarding data through techniques like differential privacy, which was not applied here.

Answer

Fairness

Answer

Transparency

Answer

Accountability

Question 12

What is the bias-variance tradeoff in machine learning?

Accepted Answer

The balance between model simplicity (underfitting) and model complexity (overfitting). The bias-variance tradeoff directly addresses the tension between underfitting (high bias, overly simple model) and overfitting (high variance, overly complex model). In Azure Machine Learning, this tradeoff is managed through hyperparameter tuning (e.g., regularization strength, tree depth) to achieve optimal generalization on unseen data.

Answer

Choosing between model accuracy and computational cost

Answer

Deciding whether to use biased training data or unbiased test data

Answer

The tradeoff between training speed and model size

Question 13

A robotic arm in a factory needs to pick parts from a bin. The system must identify each part and its exact outline to ensure precise grasping. Which Computer Vision capability should be used?

Accepted Answer

Semantic segmentation. Semantic segmentation is the correct capability because it classifies each pixel in an image, providing a precise outline of each part. This pixel-level classification is essential for a robotic arm to determine the exact shape and boundaries of parts for accurate grasping, unlike object detection which only provides bounding boxes.

Answer

Object detection

Answer

Image classification

Answer

Optical Character Recognition

Question 14

A bank uses a machine learning model to predict credit card fraud. The model's output is a probability score. The business wants to minimize the number of false positives (legitimate transactions incorrectly flagged as fraud) because these cause customer dissatisfaction. At the same time, they must also catch most fraudulent transactions. Which metric should the bank optimize to balance these two goals?

Accepted Answer

D: F1 score. The F1 score is the harmonic mean of precision and recall, making it the ideal metric when a balance between minimizing false positives (precision) and catching most fraudulent transactions (recall) is required. In this credit card fraud detection scenario, optimizing F1 ensures the model reduces customer dissatisfaction from false positives while still maintaining high detection of actual fraud.

Answer

A: Accuracy

Answer

B: Precision

Answer

C: Recall

Question 15

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 16

A legal firm needs to automatically process thousands of court documents. The system must identify and redact sensitive personal information such as names, addresses, and social security numbers. Additionally, it must extract legal-specific entities like case numbers, judge names, and statute references. The firm has a small set of manually annotated documents with these legal entities. Which combination of Azure AI Language features should they use?

Accepted Answer

PII detection and custom named entity recognition (NER). The firm needs both PII detection to redact sensitive personal information and custom NER to extract legal-specific entities like case numbers and judge names from a small set of annotated documents. Azure AI Language provides a prebuilt PII detection feature for common sensitive data and a custom NER capability that can be trained on the firm's annotated documents to recognize domain-specific entities.

Answer

Prebuilt entity recognition and key phrase extraction

Answer

Sentiment analysis and language detection

Answer

PII detection only

Question 17

A wildlife research team uses drone imagery to monitor penguin populations in a remote area. The penguins are small, blend into the rocky background, and are often only partially visible. The team has a limited set of 500 labeled drone images showing penguins. They want to build a system that accurately detects and counts penguins. Which approach should they take using Azure AI services?

Accepted Answer

Train a Custom Vision object detection model using the labeled images.. The pre-built Computer Vision object detection API is optimized for common objects and may not perform well on small, camouflaged penguins in rocky terrain. Custom Vision allows the team to train a dedicated object detection model using their 500 labeled images, enabling the model to learn the specific visual features of penguins in this challenging environment. This approach is ideal for domain-specific detection tasks where off-the-shelf models lack accuracy.

Answer

Use the pre-built Computer Vision object detection API directly.

Answer

Use the Computer Vision Image Analysis API with the 'dense captioning' feature.

Answer

Train a Custom Vision image classification model with the labeled images.

Question 18

A data scientist is using Azure Automated Machine Learning to build a binary classification model for a highly imbalanced dataset (95% negative, 5% positive). The data scientist wants AutoML to select the best model based on a metric that is robust to class imbalance. Which primary metric should the data scientist configure in the AutoML settings?

Accepted Answer

AUC_weighted. AUC_weighted is the correct primary metric for imbalanced binary classification because it computes the area under the ROC curve for each class and averages them with weight proportional to the class support. This weighting ensures that the metric reflects performance on both the majority (95% negative) and minority (5% positive) classes, making it robust to severe imbalance. Azure Automated Machine Learning uses AUC_weighted as a recommended metric when the dataset is skewed, as it penalizes models that ignore the minority class.

Answer

Accuracy

Answer

F1_score

Answer

Log_loss

Question 19

A legal research firm uses Azure OpenAI Service to answer questions about specific case law documents. They want the model to base its answers exclusively on the content of the provided documents, without using any external knowledge from its training. Which approach should they use?

Accepted Answer

Use the 'Azure OpenAI on your data' feature with a 'Search' data source containing the documents. The 'Azure OpenAI on your data' feature with a 'Search' data source allows the model to retrieve and ground its answers exclusively on the content of the provided documents. This approach uses a search index (e.g., Azure Cognitive Search) to fetch relevant document chunks and inject them into the prompt, ensuring the model does not rely on its pre-trained knowledge. It is the only method that enforces strict document-based grounding without external knowledge leakage.

Answer

Increase the 'temperature' parameter to 0.0

Answer

Use the system message to instruct the model to only use provided documents

Answer

Set the 'max_tokens' parameter to a low value

Question 20

A data scientist is evaluating a binary classification model that predicts whether a transaction is fraudulent. The test set contains 1,000 transactions: 990 legitimate and 10 fraudulent. The model's predictions are shown in the confusion matrix below. Confusion matrix: Predicted Legitimate Predicted Fraudulent Actual Legitimate 942 48 Actual Fraudulent 2 8 Which metric should the data scientist prioritize if the business goal is to minimize the number of fraudulent transactions that are missed (false negatives)?

Accepted Answer

Recall. Recall (sensitivity) measures the proportion of actual positives correctly identified, calculated as TP/(TP+FN). With 2 false negatives (missed fraudulent transactions), recall is 8/(8+2)=0.80. Minimizing missed fraud directly corresponds to maximizing recall, making it the correct priority for this business goal.

Answer

Precision

Answer

Accuracy

Answer

Specificity