Microsoft Azure AI Fundamentals AI-900 AI-900 Questions 976–985 | Page 14/14

976

MCQeasy

Which of the following is an example of 'anomaly detection' as an AI workload?

A.Translating customer support emails from Spanish to English

B.Automatically identifying fraudulent credit card transactions that deviate from a customer's normal patterns

C.Generating product descriptions from a list of specifications

D.Classifying customer reviews as positive or negative

AnswerB

Anomaly detection models learn a statistical baseline of a customer's typical spending behaviour — such as amounts, merchant categories, times, and locations — and flag transactions that are highly improbable under that baseline. Because fraud patterns are often novel and unlabelled, this is a classic unsupervised anomaly-detection use case, not a supervised classification task. That is why identifying credit card transactions that deviate from a customer's normal patterns is the correct answer.

Why this answer

Anomaly detection identifies data points that deviate significantly from the norm. In this case, fraudulent credit card transactions are detected because they do not match the customer's typical spending patterns, which is a classic use case for anomaly detection in AI workloads.

Exam trap

The trap here is that candidates may confuse anomaly detection with classification (Option D) because both involve identifying unusual items, but classification requires labeled training data for known categories, whereas anomaly detection focuses on deviations from a learned norm without predefined labels for anomalies.

How to eliminate wrong answers

Option A is wrong because translating emails from Spanish to English is a natural language processing (NLP) task for machine translation, not anomaly detection. Option C is wrong because generating product descriptions from specifications is a generative AI or natural language generation task, not anomaly detection. Option D is wrong because classifying customer reviews as positive or negative is a text classification or sentiment analysis task, which falls under supervised learning, not anomaly detection.

Full explanation →

977

MCQeasy

A bank deploys an AI system to automatically approve or reject loan applications. After six months, an audit reveals that the system approves loans at a significantly lower rate for applicants from a specific ethnic group compared to other groups with similar financial profiles. Which Microsoft responsible AI principle is most directly violated by this outcome?

A.Transparency

B.Fairness

C.Privacy

D.Reliability

AnswerB

Fairness requires that AI systems treat all people fairly and do not discriminate based on sensitive attributes like ethnicity. The significantly lower approval rate for one ethnic group despite similar financial profiles is a direct violation of Fairness.

Why this answer

The AI system's approval rate disparity for a specific ethnic group, despite similar financial profiles, directly violates the Fairness principle. Fairness requires that AI systems treat all groups equitably and avoid discrimination based on sensitive attributes like ethnicity. This outcome demonstrates a lack of fairness in the model's decision-making process.

Exam trap

The trap here is that candidates may confuse the discriminatory outcome (a Fairness issue) with a lack of Transparency, thinking that if the system were more explainable the bias would be avoided, but the core violation is the unequal treatment itself.

Why the other options are wrong

The outcome describes a disparity in loan approval rates across ethnic groups, which directly violates the fairness principle, not transparency. Transparency concerns explainability and disclosure of system behavior, not the presence of bias.

The question describes disparate approval rates across ethnic groups, which directly violates fairness, not privacy. Privacy concerns data protection and consent, not discriminatory outcomes.

The question describes a bias in loan approvals across ethnic groups, which directly violates the Fairness principle. Reliability concerns system performance and accuracy, not disparate impact on protected groups.

When would these options actually be correct?

Transparency would be correct if the question described a system whose decisions could not be explained or documented, e.g., 'An AI system denies loans but cannot provide reasons for its decisions, and the bank cannot explain how the model works.'

An AI system that exposes applicants' personal financial data without consent or leaks sensitive information would violate privacy. For example, a loan approval system that stores or shares applicant data insecurely would make privacy the correct answer.

An AI system for medical diagnosis frequently produces incorrect results due to data drift, leading to misdiagnoses. In that scenario, the Reliability principle would be most directly violated.

Why candidates pick the wrong answer

Candidates may confuse 'unfair outcome' with 'lack of transparency' because they assume that if the system were transparent, the bias would be visible and thus avoided. However, transparency alone does not prevent bias.

Candidates may confuse fairness with privacy because both involve ethical handling of personal data, but privacy focuses on data protection while fairness focuses on equitable treatment.

Candidates may confuse 'unfair outcomes' with 'unreliable system,' thinking that biased results indicate the system is not working reliably, but reliability focuses on consistency and accuracy, not fairness across groups.

Full explanation →

978

MCQeasy

A data scientist trains a classification model to distinguish between images of cats and dogs. The model achieves 99% accuracy on the training set but only 75% accuracy on a validation set. Which concept best describes this situation?

A.Underfitting

B.Overfitting

C.Model bias

D.Data leakage

AnswerB

Overfitting is the correct diagnosis. The model achieves high accuracy on its training data because it has effectively memorized the training examples, including their random noise and idiosyncrasies, rather than learning the underlying pattern. When confronted with unseen validation data, that memorized detail does not transfer, so validation performance drops sharply. This high-variance behavior is the signature of too much model capacity relative to the amount of training data.

Why this answer

The model performs exceptionally well on the training data (99% accuracy) but significantly worse on unseen validation data (75% accuracy). This gap indicates the model has memorized noise and specific patterns in the training set rather than learning generalizable features, which is the classic definition of overfitting.

Exam trap

The trap here is that candidates see high accuracy and assume the model is good, failing to recognize that the large gap between training and validation accuracy is the hallmark of overfitting, not underfitting or bias.

Why the other options are wrong

Underfitting occurs when a model performs poorly on both training and validation sets, but here the model achieves 99% accuracy on training data, indicating it has learned the training data too well, not too little.

Model bias refers to errors due to overly simplistic assumptions, but here the model performs well on training data (99%) and poorly on validation (75%), indicating overfitting, not bias.

Data leakage refers to the model using information not available at inference time, which typically leads to overly optimistic performance. In this case, the model performs worse on validation data, not better, so leakage is not indicated.

When would these options actually be correct?

A scenario where a model has low accuracy on both training and validation sets (e.g., 60% on training and 58% on validation) for a classification task, indicating it is too simple to capture underlying patterns.

Model bias would be correct if the model had low accuracy on both training and validation sets (e.g., 60% on both), suggesting it fails to capture underlying patterns due to assumptions like linearity in a nonlinear problem.

A scenario where a model achieves near-perfect accuracy on both training and validation sets, but fails in production because the validation data inadvertently contained information from the future or from the target variable (e.g., using patient ID to predict disease).

Why candidates pick the wrong answer

Candidates may confuse high training accuracy with good generalization, not realizing that underfitting is characterized by poor performance on training data, not just validation data.

Candidates may confuse 'bias' with any performance discrepancy, not realizing that bias typically causes poor performance across all datasets, not just validation.

Candidates may confuse high training accuracy with leakage, thinking that the model 'cheated' by learning patterns that don't generalize, but leakage usually inflates validation accuracy as well.

Full explanation →

979

MCQmedium

What is precision in the context of binary classification model evaluation?

A.The proportion of actual positives that the model correctly identified

B.The proportion of positive predictions that are actually correct

C.The overall proportion of all predictions that are correct

D.The number of decimal places in the model's confidence score

AnswerB

Precision = true positives / (true positives + false positives). It measures how many of the model's positive predictions were actually correct, making it a direct check on false alarms. High precision means that when the model says a result is positive, it is usually right.

Why this answer

Precision measures the accuracy of positive predictions: it is the ratio of true positives to the sum of true positives and false positives. Option B correctly defines this as 'the proportion of positive predictions that are actually correct,' which is the standard definition used in Azure Machine Learning's classification metrics.

Exam trap

The trap here is that candidates often confuse precision with recall (Option A) because both involve true positives, but precision focuses on the correctness of positive predictions while recall focuses on capturing all actual positives.

How to eliminate wrong answers

Option A is wrong because it describes recall (sensitivity), not precision; recall focuses on how many actual positives were caught, not how many predicted positives were correct. Option C is wrong because it describes accuracy, which is the overall proportion of correct predictions (both positives and negatives) out of total predictions, not precision. Option D is wrong because it confuses the mathematical concept of precision in classification with numerical precision (decimal places) in confidence scores, which is unrelated to model evaluation metrics.

Full explanation →

980

MCQmedium

A hospital deploys an AI system to assist doctors in interpreting MRI scans. The system highlights the regions of interest and provides a numeric confidence score for its findings, along with a list of the image features that contributed to the diagnosis. Which responsible AI principle is being applied?

A.Fairness

B.Transparency

C.Privacy

D.Accountability

AnswerB

Transparency is the correct principle because it refers to the degree to which an AI system’s decision-making process can be understood by humans. In this clinical scenario, the AI system assists doctors by providing interpretable outputs—for example, highlighting which patient features like vital signs or lab results influenced a diagnosis. This aligns with Microsoft’s responsible AI principle of transparency, which enables clinicians to validate AI suggestions and build trust in the technology. Without such explanations, doctors could not meaningfully assess the reliability of the AI's recommendations.

Why this answer

The system provides a numeric confidence score and a list of image features that contributed to the diagnosis, which directly supports the principle of Transparency. Transparency in responsible AI requires that AI systems are understandable and that their decisions can be explained to users, enabling clinicians to interpret and trust the output.

Exam trap

The trap here is that candidates confuse Transparency with Accountability, thinking that providing a confidence score implies responsibility, but Transparency is specifically about making the model's reasoning visible and interpretable to users.

Why the other options are wrong

The system's focus on highlighting regions, providing confidence scores, and listing contributing features directly addresses transparency (explainability), not fairness. Fairness would involve ensuring the model performs equitably across demographic groups, which is not described.

The system highlights regions of interest, provides confidence scores, and explains image features, which directly addresses transparency (explainability), not privacy. Privacy concerns data protection and consent, which are not mentioned in the scenario.

Accountability refers to assigning responsibility for AI outcomes, but the question describes the system explaining its reasoning (features and confidence), which is about transparency, not accountability.

When would these options actually be correct?

A hospital deploys an AI system to prioritize patients for organ transplants based on medical history and socioeconomic data. The system is found to allocate fewer organs to patients from certain racial backgrounds. Which principle is being violated?

A hospital deploys an AI system that analyzes patient MRI scans and stores the images and diagnoses in a cloud database without anonymizing patient identifiers. Which responsible AI principle is being violated?

A healthcare AI system makes a diagnostic error, and the hospital needs to determine who is responsible—the developer, the hospital, or the clinician. The principle of accountability would be applied to ensure clear ownership and remediation processes.

Why candidates pick the wrong answer

Candidates may confuse the general ethical importance of fairness with the specific scenario, or assume that any responsible AI principle must include fairness, overlooking that the question explicitly describes explainability features.

Candidates may confuse the handling of sensitive medical data with the principle of privacy, assuming any AI system in healthcare must prioritize privacy, even when the scenario focuses on explainability.

Candidates may confuse accountability with transparency because both involve oversight, but accountability focuses on responsibility for outcomes, not on explaining how decisions are made.

Full explanation →

981

MCQmedium

A company wants to use Azure OpenAI to generate realistic customer conversations for training a chatbot. They have a set of example conversation snippets and want the model to mimic the style and structure of those examples. The company does not want to retrain the model. Which approach should they use?

A.Fine-tune the model on the conversation dataset

B.Use prompt engineering with few-shot examples in the prompt

C.Use DALL-E to generate the conversations

D.Apply a content filter to restrict the output style

AnswerB

Few-shot prompting provides a small set of example conversations directly in the prompt, enabling the model to infer the desired style, tone, and format through in-context learning. The model then generates new realistic customer conversations that mirror those examples, without any weight updates or retraining. This approach is efficient, flexible, and preserves the original model, making it the best choice here.

Why this answer

Prompt engineering with few-shot examples allows the model to mimic the style and structure of provided conversation snippets without retraining. By including a few example conversations in the prompt, the model learns the desired pattern through in-context learning, leveraging its pre-trained capabilities to generate realistic customer conversations.

Exam trap

The trap here is that candidates may confuse fine-tuning with in-context learning, assuming that any style adaptation requires retraining, when in fact few-shot prompting can achieve the same result without modifying the model.

How to eliminate wrong answers

Option A is wrong because fine-tuning requires retraining the model on the conversation dataset, which contradicts the requirement that the company does not want to retrain the model. Option C is wrong because DALL-E is designed for image generation, not text-based conversation generation, and cannot produce realistic customer conversations. Option D is wrong because content filters restrict output based on safety or policy rules, but they do not control or mimic the style and structure of example conversations.

Full explanation →

982

MCQmedium

A data scientist is training a model to predict whether a customer will purchase a product (Yes/No). The dataset contains 90% 'No' and 10% 'Yes'. After training, the model achieves 90% accuracy. Which evaluation metric would be more informative to assess the model's performance on the minority class?

A.Mean Absolute Error (MAE)

B.F1 score

C.Area Under the Curve (AUC)

D.R-squared

AnswerB

F1 score is the harmonic mean of precision and recall, giving equal weight to both metrics. It is particularly valuable when the dataset is imbalanced because it directly penalizes false positives and false negatives from the minority-class perspective. Unlike accuracy, F1 does not let a majority-class bias dominate the evaluation, as it only considers the positive class's precision and recall. Therefore, it provides a single, interpretable number that reflects how well the model detects the rare (positive) outcome while maintaining reliability of its positive predictions.

Why this answer

In this imbalanced dataset (90% 'No', 10% 'Yes'), a model that always predicts 'No' would achieve 90% accuracy, making accuracy a misleading metric. The F1 score is the harmonic mean of precision and recall, specifically designed to evaluate a model's performance on the minority class by balancing false positives and false negatives. It is the most informative metric here because it directly measures how well the model identifies the rare 'Yes' purchases without being inflated by the majority class.

Exam trap

The trap here is that candidates see 90% accuracy and assume the model is performing well, failing to recognize that accuracy is misleading in imbalanced datasets, and they may incorrectly select AUC because it is a common classification metric, but it does not directly penalize poor minority-class performance like the F1 score does.

Why the other options are wrong

MAE is a regression metric that measures average absolute error between continuous predictions and actual values, not suitable for binary classification tasks like predicting Yes/No.

AUC measures the model's ability to distinguish between classes across all thresholds, but it does not specifically focus on the minority class's precision and recall, which are critical for imbalanced datasets.

R-squared is a metric for regression models, measuring the proportion of variance explained, but this is a binary classification problem (Yes/No purchase).

When would these options actually be correct?

When the question involves predicting a continuous value, such as house price or temperature, and the evaluation requires a measure of average prediction error in the original units.

AUC would be correct in a question asking for a metric that evaluates the overall ranking performance of a binary classifier, independent of threshold choice, such as comparing different models' discriminative power on a balanced dataset.

R-squared would be correct for a regression question, e.g., 'A model predicts house prices. Which metric evaluates how well the model explains variance in the target variable?'

Why candidates pick the wrong answer

Candidates may confuse MAE with classification error rate or think it applies to any prediction task, overlooking that it is designed for regression, not binary outcomes.

Candidates may think AUC is always the best metric for imbalanced data because it is threshold-independent, but it does not directly assess performance on the minority class like F1 does.

Candidates may confuse R-squared as a general performance metric applicable to any model, not realizing it is specific to regression.

Full explanation →

983

MCQhard

What is 'pose estimation' in computer vision and what is it used for?

A.Estimating the correct posture for employees based on ergonomics guidelines

B.Detecting body keypoint positions (joints) in images to infer posture and movement

C.Determining the camera angle and position used to capture a photograph

D.Classifying whether a person is sitting or standing in an image

AnswerB

Pose estimation is a computer vision technique that uses neural network regression or heatmap-based decoding to predict the 2D or 3D coordinates of body joints (keypoints) from an image. These keypoints, often following a skeletal model like COCO's 17-joint layout, allow algorithms to reconstruct the subject's posture and track movement across frames. This output underpins applications such as fitness coaching, motion capture, and gesture recognition.

Why this answer

Pose estimation is a computer vision technique that detects and localizes keypoints (joints) on a human body in an image or video. These keypoints, such as shoulders, elbows, wrists, hips, and knees, are used to infer the body's posture, orientation, and movement. Option B correctly describes this process of detecting body keypoint positions to infer posture and movement.

Exam trap

The trap here is confusing human pose estimation (detecting body keypoints) with camera pose estimation (determining camera position) or with simple classification tasks like sitting/standing, leading candidates to pick options C or D.

How to eliminate wrong answers

Option A is wrong because it describes an ergonomic assessment, not a computer vision technique; pose estimation outputs keypoint coordinates, not compliance with ergonomic guidelines. Option C is wrong because determining camera angle and position is a separate task called camera pose estimation or structure from motion, not human pose estimation. Option D is wrong because classifying a person as sitting or standing is a simpler action recognition task that could be derived from pose estimation, but pose estimation itself involves detecting specific joint keypoints, not just outputting a binary state.

Full explanation →

984

MCQmedium

A financial services company uses an AI system to detect fraudulent credit card transactions. After deployment, the system incorrectly flags a significant number of legitimate transactions as fraudulent, causing customer dissatisfaction. The company wants to reduce these false positives while still catching most fraudulent transactions. Which Microsoft responsible AI principle should guide their redesign of the system?

A.Reliability and safety

B.Fairness

C.Transparency

D.Privacy and security

AnswerA

In fraud detection, reliability hinges on tuning the decision threshold to balance false positives and false negatives—too many false positives block legitimate transactions, while false negatives let fraudulent activity slip through. Safety extends this by demanding robust performance under real-world data drift and graceful handling of edge cases, so the system's errors remain bounded and non-harmful. This principle directly governs prediction accuracy and error trade-offs.

Why this answer

The Reliability and safety principle emphasizes that AI systems should perform reliably, safely, and consistently under normal conditions. In this scenario, the high rate of false positives indicates the system is not operating reliably for legitimate transactions, causing customer harm. Redesigning to reduce false positives while maintaining fraud detection aligns directly with improving the system's reliability and safety for end users.

Exam trap

The trap here is that candidates confuse 'false positives causing customer dissatisfaction' with a fairness or transparency issue, when in fact it is a reliability and safety problem about the system's accuracy and trustworthiness in production.

How to eliminate wrong answers

Option B (Fairness) is wrong because the issue is not about bias or discrimination against protected groups—false positives affect all legitimate customers equally, not a specific demographic. Option C (Transparency) is wrong because the problem is not about explaining how decisions are made; customers are dissatisfied due to incorrect flags, not a lack of explanation. Option D (Privacy and security) is wrong because the system is not leaking or mishandling personal data; the core issue is classification accuracy, not data protection.

Full explanation →

985

MCQmedium

What is the difference between Azure OpenAI Service and the public OpenAI API?

A.Azure OpenAI has different models that perform better than OpenAI

B.Azure OpenAI adds enterprise security, compliance, private networking, and Azure integration

C.Azure OpenAI is only available for government customers

D.Azure OpenAI does not support GPT-4 models

AnswerB

Azure OpenAI delivers the same OpenAI model weights with enterprise-grade additions: private connectivity via Azure Virtual Network and private endpoints, role-based access control through Azure RBAC, and integration with services like Azure Cognitive Search and Azure Logic Apps. Customer data is not used to train models, and the service offers compliance certifications such as SOC 2, ISO 27001, HIPAA, and FedRAMP. This enterprise layer makes Azure OpenAI suitable for organizations needing strict data governance, isolation, and management.

Why this answer

Azure OpenAI Service is a Microsoft Azure-based offering that wraps the same underlying OpenAI models (GPT-4, GPT-3.5, etc.) with enterprise-grade features such as Azure Active Directory authentication, private endpoints via Azure Virtual Network, compliance certifications (e.g., ISO 27001, SOC 2), and seamless integration with other Azure services like Cognitive Search and Logic Apps. The public OpenAI API lacks these enterprise controls, making Azure OpenAI the preferred choice for organizations that require data residency, network isolation, and managed identity access.

Exam trap

The trap here is that candidates assume 'Azure OpenAI' is a completely different set of models or a restricted service, when in fact it is the same models with added enterprise security and integration features.

How to eliminate wrong answers

Option A is wrong because Azure OpenAI Service uses the same underlying models as the public OpenAI API (e.g., GPT-4, GPT-3.5, Codex); there are no 'different models that perform better' — performance differences arise from deployment configuration, not model variants. Option C is wrong because Azure OpenAI Service is available to all Azure customers globally, not exclusively to government customers (though a separate Azure Government instance exists for US government agencies). Option D is wrong because Azure OpenAI Service fully supports GPT-4 models, including GPT-4 Turbo and GPT-4o, with the same capabilities as the public API.

Full explanation →

Microsoft Azure AI Fundamentals AI-900 (AI-900) — Questions 976–985