Scenario practice questions

Inclusiveness

Why wrong: Inclusiveness focuses on empowering a diverse range of people, including users with disabilities, and ensuring that the AI system is accessible and built with input from varied communities. The concern in this scenario is not about who is included in the design process or the system's accessibility, but about whether the system's analysis treats different categories of vehicles equitably. That distinction places it under Fairness rather than Inclusiveness.

Reliability and safety

Why wrong: Reliability and safety address whether the AI system functions consistently, is robust to errors, and avoids causing physical or operational harm in its given environment. The scenario does not describe system crashes, inaccurate detections, or unsafe actions; it describes a pattern of unequal treatment across vehicle classes. A system could be perfectly reliable and safe while still being unfair, so this principle does not address the core issue.

Transparency

Why wrong: Transparency pertains to the visibility and interpretability of an AI system's workings, such as providing explanations for why a particular decision was made or making the model's logic open to review. The problem in the scenario is not that the system's logic is hidden, but that the consequences of its analysis are inequitable across vehicle types. Without a fairness evaluation, explaining the system's decisions would not remedy the underlying bias.

Conversational Language Understanding (CLU)

Why wrong: Conversational Language Understanding (CLU) predicts the user's intent (like 'cancel_order') and extracts relevant entities (like order ID) to support a dialogue flow or trigger a downstream action. It does not store question-answer pairs from a document, and while it can be orchestrated to route to a QnA service, CLU alone cannot fetch an answer and present it. Its output is a structured schema for decision-making, not an answer string, so it is not the correct primary service for knowledge-base question answering.

Custom Text Classification

Why wrong: Custom Text Classification trains a machine learning model to assign one or more predefined labels, such as 'billing' or 'technical issue', to a user's input text based on your labeled training data. After training, the model only outputs label predictions; it cannot index a FAQ document, match a question to an answer, or return evidence from source content. This makes it appropriate for routing or triage, but not for directly answering customer questions with content pulled from a knowledge base.

Language Detection

Why wrong: Language Detection is a text analytics capability that classifies a single input string by its primary written language, returning an ISO language code and a confidence score. It performs no retrieval, no knowledge-base lookup, and no extraction of answer content; it just answers the question 'what language is this?' Even if you detected or translated the customer's question, it would never produce the factual answer from a document. Therefore, while useful for multilingual routing, it cannot power a Q&A chatbot.

Recognizing counterfeit products in supply chain images

Why wrong: Counterfeit detection requires verifying whether an item is authentic, which involves subtle visual anomalies and provenance data that demand specialized fine-tuning and custom models per brand. Product recognition only determines that an item looks like a known catalog product; it does not authenticate legitimacy or detect imitation manufacturing, so it is insufficient for this anti-counterfeiting task.

Recognizing products mentioned in customer text reviews

Why wrong: Extracting product names from text reviews requires natural language processing (NLP), such as named entity recognition with Azure AI Language, to parse unstructured customer feedback. The product recognition capability analyzed in this question is strictly a computer vision service that accepts image inputs and never processes text, so applying it to reviews is a modality mismatch.

Detecting product defects in manufacturing quality control

Why wrong: Defect detection is a specialized visual inspection task that identifies anomalies, scratches, or structural flaws on manufactured items, typically using custom-trained models on defect datasets. Azure AI Vision's product recognition service, by contrast, is a prebuilt retail solution that classifies whole, undamaged products into a known catalog, not a quality-assurance system.

Regression

Why wrong: Regression is inappropriate because it outputs a continuous numeric value, such as a transaction amount or a real-valued risk score. The fraud-detection objective is to assign a categorical label, not to estimate a quantity on an unbounded scale. While some classifiers output probabilities, those probabilities are then thresholds to produce a class, which is still a classification workflow. Predicting a continuous outcome directly would not map cleanly to the two desired categories.

Multi-class classification

Why wrong: Multi-class classification would be required if the target variable had three or more distinct categories, such as fraud type A, type B, and legitimate. Here the prediction task involves only a binary outcome, so the additional complexity of multi-class algorithms is unnecessary. Even if a model like a neural network with a softmax output could technically handle two classes, standard binary classification is the standard and most efficient framing. The number of mutually exclusive classes is exactly two, not more.

Clustering

Why wrong: Clustering is an unsupervised learning technique that identifies natural groupings within unlabeled data based on feature similarity. The fraud-detection dataset includes labeled examples, so the correct approach is supervised classification that leverages these labels to establish a mapping from features to outcomes. Clustering would ignore the ground-truth labels and would instead partition transactions into clusters that may or may not align with fraud status. Without labels, there is no way to guide the model toward the specific fraudulent versus legitimate decision boundary.

Object detection

Why wrong: Object detection can detect people in frames, but it does not track their movement across frames or count them entering/leaving specific zones.

Optical Character Recognition (OCR)

Why wrong: OCR extracts text from images and is unrelated to detecting people or tracking movement.

Semantic segmentation

Why wrong: Semantic segmentation classifies each pixel into categories (e.g., person, floor), but it does not track individuals or count zone entries over time.

Transparency

Why wrong: Transparency aims to make AI decisions interpretable and explainable to stakeholders, such as providing clear justifications for why a traffic light timing is adjusted. However, a fully transparent system could still exhibit unequal accuracy across neighborhoods; transparency alone does not prevent or measure performance disparities. The core issue here is the imbalance in prediction quality, which transparency does not directly tackle.

Accountability

Why wrong: Accountability deals with assigning responsibility for AI outcomes and establishing governance frameworks for auditing and remediation. It is a process-level concern about who is responsible, not a property of the model's performance itself. The fact that accuracy differs among neighborhoods is a substantive fairness flaw, not merely a question of accountability.

Privacy and security

Why wrong: Privacy and security concerns safeguarding data from unauthorized access, ensuring encryption, access controls, and compliance with regulations like GDPR. It does not address the issue of model accuracy varying across neighborhoods, which is about algorithmic bias. Even with perfect data protection, the AI system could still produce inaccurate results for certain groups, so this option is incorrect.

Real-time transcription of live audio streams for immediate use

Why wrong: Batch transcription is designed for pre-recorded audio files submitted in bulk, processed asynchronously, with results retrieved later via status polling. Real-time transcription streams live audio through a low-latency API (e.g., WebSocket) to produce immediate captions or subtitles; this is a distinct feature and not what batch transcription does. Batch mode lacks the real-time return of word-by-word results.

Synchronizing speech transcription across multiple languages simultaneously

Why wrong: Batch transcription transcribes audio per-file using a specified source language locale; it does not synchronously produce multi-language transcripts or translations. Multi-language simultaneous transcription or translation is handled by separate Azure capabilities like speech translation or multilingual conversation transcription, which address language coverage rather than bulk asynchronous volume. Selecting this option mistakenly equates 'batch' with 'multilingual,' but batch refers to processing scale, not language diversity.

Training a custom speech recognition model on audio samples

Why wrong: Batch transcription is an asynchronous service that applies the Speech service's existing recognition models to pre-recorded audio files; it does not train or adapt models. Custom speech model training is a separate capability where you upload labeled audio or text to build a personalized acoustic/language model, then deploy that model for recognition. Choosing this option confuses model creation with bulk batch processing.

Reliability and Safety

Why wrong: Reliability and Safety concern whether the AI system performs its intended function consistently and without causing physical, financial, or operational harm due to errors, crashes, or unexpected behavior. While a biased lending model could be seen as unreliable, the core issue here is not a malfunction or safety failure but rather the learned discriminatory patterns from training data. The system may be highly reliable in terms of uptime and accuracy, yet still produce unfair results, which is a distinct problem.

Privacy and Security

Why wrong: Privacy and Security involve protecting personal and financial data from unauthorized access, leakage, or misuse through encryption, access controls, and compliance with data protection regulations. The loan recommendation bias described is not a data breach or confidentiality failure; the system may fully protect user data yet still render unfair decisions based on biased historical labels. Therefore, while privacy and security are critical, they address a different risk than the fairness violation at issue here.

Inclusiveness

Why wrong: Inclusiveness focuses on designing AI systems that are accessible and usable by people of all abilities, backgrounds, and languages, such as supporting screen readers, providing multilingual interfaces, or accommodating diverse user needs. The bias in loan recommendations does not stem from a lack of accessibility or exclusion from use, but rather from disparate treatment of protected classes in the decision-making logic. Thus, improving inclusiveness would not directly remedy the discriminatory outcomes described.

Overfitting

Why wrong: Overfitting is characterized by a model that memorizes training data, including its noise, so training loss is typically very low while test loss is high. This creates a wide performance gap, not uniformly high error on both splits. Since the scenario describes high error on training and test sets alike, overfitting cannot explain the pattern; the model would at least succeed on the data it was fitted to.

Data leakage

Why wrong: Data leakage happens when information from the target variable or from the test set is unintentionally used during training, such as through target-encoded features or improper preprocessing applied before splitting. Leakage typically inflates model evaluation metrics and makes test performance look unrealistically good, not poor. Therefore, it would produce the opposite symptom—abnormally high accuracy—rather than high error on both training and test data.

Feature scaling

Why wrong: Feature scaling, such as standardization or normalization, is a preprocessing technique that improves optimization convergence and ensures equal influence of feature magnitudes, especially for gradient-based models or distance-based algorithms. Omitting feature scaling can slow training or cause poor convergence in extreme cases, but it does not inherently cause high error across both training and test sets. The model can still learn the pattern with enough iterations, so scaling is not a fundamental cause of underfitting.

Fairness

Why wrong: A fairness violation would require evidence that the model's screening decisions systematically disadvantage protected groups, such as by rejecting candidates of a particular ethnicity at a higher rate. The scenario instead centers on the operator's inability to articulate why a specific applicant was rejected, which is an explainability concern. Without access to the decision logic, one cannot even determine whether bias exists, so fairness is not the immediate principle breached.

Accountability

Why wrong: Accountability is about establishing which person or team owns the model and its outcomes, including maintaining audit trails and taking corrective action. Although an AI developer or HR manager should ultimately be answerable for the resume screener, the described failure is a lack of explanation for the model's judgment, not a lack of ownership. Thus, accountability is relevant to governance but does not pinpoint the transparency deficiency at issue.

Privacy and security

Why wrong: Privacy and security govern how resume data is collected, stored, accessed, and protected from breaches or unauthorized use. Nothing in the scenario suggests that candidate information was exposed, misused, or handled without consent, so a data-protection violation is not present. The central problem remains the inability to explain a decision, which is unrelated to confidentiality or infrastructure safeguards.

Fairness

Why wrong: Fairness is not the applicable principle here because the scenario centers on physical comfort and operational integrity, not on equitable treatment across demographic groups. Fairness would apply if the AI allocated resources or made decisions that could systematically disadvantage certain people, such as varying temperature setpoints based on tenants' characteristics. Since the prompt describes a generic HVAC control problem without any bias or discrimination dimension, fairness is a plausible-sounding but incorrect choice.

Transparency

Why wrong: Transparency is incorrect because it concerns the explainability and interpretability of AI decisions — how users or auditors can understand why a model made a particular output. In an HVAC system, even if the AI's scheduling logic were fully interpretable, transparency alone would not prevent uncomfortable temperature swings or mechanical failures. The question is about the system's operational trustworthiness, not about providing explanations to building occupants or facility managers.

Privacy and security

Why wrong: Privacy and security is wrong because no personal data or sensitive information is mentioned in the scenario; the system monitors environmental variables like temperature and occupancy, not individuals' identities. Security would matter if there were data to protect or if adversarial threats could compromise the system, but the stated issue is the physical safety and comfort impact of incorrect adjustments. Protecting personal data does not inherently ensure that an HVAC system operates reliably and avoids causing discomfort.

Fairness

Why wrong: Fairness in responsible AI is concerned with ensuring outcomes and decisions do not systematically disadvantage protected groups; it is evaluated through group-level metrics such as demographic parity or equalized odds. Re-identifying individuals in a medical dataset is a privacy breach that violates confidentiality, not a fairness failure. Even a model that treats all demographic groups equitably can still expose patient identities, so this option misidentifies the ethical principle at stake.

Transparency

Why wrong: Transparency refers to clearly communicating an AI system's capabilities, limitations, and decision-making logic to stakeholders, ensuring users understand when and how AI is used. It does not speak to whether underlying data can be re-identified; a fully transparent system could still leak patient identities. This option is wrong because the failure here is data protection, not a lack of explainability or openness about the system's behavior.

Accountability

Why wrong: Accountability in responsible AI concerns establishing clear human ownership, governance structures, and audit trails for how an AI system is developed and deployed. While a loss of patient privacy might eventually lead to accountability issues, the immediate harm described is unauthorized re-identification, which is a security and privacy failure rather than an absence of role ownership. Thus this principle does not match the specific violation in the scenario.

A spelling checker that identifies misspelled words in a document

Why wrong: A spelling checker identifies misspelled words by comparing tokens against a dictionary or language model, often using edit distance, n-gram probabilities, or neural language modeling for context. This is a rule-based or error-correction task focused on text accuracy, not on predicting a user's future choices from behavioral history. Recommendation systems solve the different problem of ranking items according to a user's inferred tastes, so the underlying purpose and ML paradigm are unrelated.

A database storing customer purchase history

Why wrong: A relational or NoSQL database that stores purchase history performs data persistence, indexing, and querying operations. Storing transactional rows is simply the ingestion layer that provides raw material for analytics. A recommendation system, by contrast, applies machine learning techniques like collaborative filtering or association rules to that historical data to infer preferences and generate personalized product suggestions. The database itself lacks any predictive logic.

A barcode scanner at a checkout counter

Why wrong: A barcode scanner uses laser or image-based decoding to read UPC/EAN codes and map them to product records. That is a perceptual, lookup-oriented computer vision task, not a predictive algorithm. While product identification can feed downstream analytics, the scanner itself does not learn from user behavior or anticipate what a customer might buy next, which is the core function of a recommendation system.

Reinforcement learning to optimize package sorting speed

Why wrong: Reinforcement learning (RL) trains an agent to make sequential decisions by maximizing cumulative rewards through trial-and-error interactions with an environment. While RL could be used to optimize conveyor-belt speeds or robot arm trajectories in a sorting facility, it does not read or interpret address text on labels. The task described is a one-step perception-and-classification problem (read address, determine destination), not a sequential policy optimization problem with delayed rewards.

Facial recognition to identify the delivery person

Why wrong: Facial recognition is a biometric computer vision technique that identifies or verifies a specific person by analyzing facial features (e.g., face landmarks, embeddings). While it could be used to authenticate a delivery person at the door, it provides no information about the package's destination address or sorting bin. The task here is text extraction and address-based routing, so facial recognition is irrelevant to the core processing pipeline.

Sentiment analysis to assess package condition

Why wrong: Sentiment analysis is a natural language processing (NLP) technique that detects emotional tone (positive, negative, neutral) in written text, typically from reviews, social media, or feedback. Package condition is a physical, non-textual attribute; there is no emotional or opinion-bearing text on a shipping label to analyze. Sorting packages by address requires reading the printed text (OCR) and then classifying or matching that text to a destination, not assessing sentiment.

Supervised learning with labeled game states

Why wrong: Supervised learning would need a dataset of game states paired with the correct action, typically obtained from expert players or solved positions, so the model could only imitate those examples. For any nontrivial game the state space is astronomically large, making comprehensive labeling impractical, and the model would fail in states outside its training distribution. Reinforcement learning, by contrast, generates its own experience and learns from reward feedback, so it can surpass human-level play without pre-labeled expert moves.

Clustering similar game strategies together

Why wrong: Clustering is an unsupervised technique that partitions data into groups based on similarity, such as grouping recorded game states by style, but it never outputs an action or learns from the consequences of moves. It has no concept of reward, win/loss, or long-term strategy, so it cannot evaluate whether a move improves the agent's chances of winning. Reinforcement learning is designed precisely for this setting because it uses the environment's reward signal to iteratively improve a policy that selects actions.

Regression to predict the final game score

Why wrong: Regression models learn a mapping from input features to a continuous numeric target, such as predicting a team's final score from historical statistics. Even a highly accurate score predictor does not produce the next action or policy needed to play the game, because gameplay requires sequential decision-making in response to an opponent. Training a game-playing agent therefore requires a framework like reinforcement learning that optimizes action choices through trial-and-error reward signals.

Azure AI Custom Vision

Why wrong: Azure AI Custom Vision is meant for creating custom image classifiers and object detectors from tagged example images, and it processes one still image at a time. It does not natively provide temporal or audio signals, so it cannot track objects as they move across video frames, identify known people in a meeting, or transcribe spoken words. Since the scenario demands video-level analysis with participant identification, object tracking, and transcription, Custom Vision is not a suitable answer.

Azure AI Face

Why wrong: Azure AI Face provides face detection, verification, and recognition on individual images, and it can return attributes like age, emotion, and head pose. It lacks the video-level capabilities of Video Indexer such as correlating faces across a shot, tracking a person throughout the recording, detecting scenes, or producing transcripts from speech. Thus, using Azure AI Face alone would fail to meet the full requirement of analyzing meeting videos for participant identification, object tracking, and transcription.

Azure AI Vision OCR

Why wrong: Azure AI Vision OCR is designed to extract printed and handwritten text from still images, not to perform temporal analysis of video. Although Video Indexer also leverages OCR to read on-screen text from video frames, the Vision OCR service alone cannot identify meeting participants, track moving objects across frames, or generate an audio transcript. Therefore, it is a component of video analysis, not a complete video-analytics solution for the described need.

Fairness

Why wrong: Fairness pertains to ensuring AI systems do not produce biased or discriminatory outcomes across groups defined by attributes such as race, gender, or age. The question's scenario contains no indication of differential treatment across demographic groups; instead, it describes a uniform failure triggered by weather conditions. Therefore, fairness is not the primary principle implicated by the autonomous vehicle's inadequate performance.

Privacy and security

Why wrong: Privacy and security focus on safeguarding sensitive data and protecting AI systems from malicious attacks, such as adversarial inputs or unauthorized access. A self-driving car failing in adverse weather is an unintended functional deficiency, not a data breach or cyberattack. Thus, while security is important for autonomous systems, it does not explain the described weather-related reliability problem.

Transparency

Why wrong: Transparency in AI refers to the degree to which an AI system's decision-making process is visible and explainable to stakeholders, such as developers or regulators. The self-driving car scenario centers on an operational failure during adverse weather, which is a matter of robust performance rather than a lack of interpretability. While transparent models may help debug failures, the core violation here is not opacity but unsafe, unreliable behavior.

A: Accuracy

Why wrong: Accuracy simply measures the proportion of correct predictions overall. In credit card fraud detection, fraudulent transactions are extremely rare—often under 0.1% of all transactions—so a model that never predicts fraud can appear 99.9% accurate while completely failing at its primary task. Because accuracy gives equal weight to false positives and false negatives, it hides the severe business cost of missed fraud and is therefore a poor performance measure for this use case.

B: Precision

Why wrong: Precision quantifies how many of the transactions flagged as fraudulent are actually fraudulent, so a high-precision model rarely harasses legitimate customers. However, achieving high precision usually requires the model to be very conservative and only flag cases with strong evidence, which allows many subtle or novel fraud patterns to go undetected. For fraud detection, optimizing only precision would sacrifice the model's ability to uncover true fraud, leaving the bank exposed to losses.

C: Recall

Why wrong: Recall measures the proportion of actual fraudulent transactions that the model successfully identifies, making it critical for catching as much fraud as possible. But a pure recall-focused approach drives the model to flag a large number of borderline legitimate transactions as suspicious, frustrating customers and flooding the bank's fraud-review queue. That excessive false-positive burden is a real operational cost, so recall in isolation is not a suitable evaluation metric.

Azure Face API

Why wrong: Azure Face API is a specialized service for detecting, analyzing, and recognizing human faces in images. It returns attributes like age, emotion, and a unique face identifier, but it performs no optical character recognition (OCR) whatsoever, so it cannot extract printed text from a shipping label.

Azure Custom Vision

Why wrong: Azure Custom Vision is a tool for training custom image classification and object detection models using labeled images. It does not have built-in OCR capabilities; it only identifies what objects or scenes are present (e.g., 'label' vs. 'package'), not the textual content on the label. To extract text, you would need to pair Custom Vision with a separate OCR service.

Azure Video Indexer

Why wrong: Azure Video Indexer is designed to analyze video and audio content, extracting insights such as spoken transcripts, faces, and on-screen text appearing in video frames. It is not intended for static images like shipping labels, and using it for single-image OCR would be inefficient and out of scope, whereas Computer Vision Read API is the direct, purpose-built solution.

Image Analysis (describing the image content)

Why wrong: Image Analysis (from Azure AI Vision) enriches an image with captions, tags, and detected objects, and can identify visual attributes such as whether content is hand-drawn or a document. However, its standard operation does not perform optical character recognition (OCR); it returns the image's semantic description rather than the textual characters present in the image. As a result, it might detect an envelope or handwritten markings but cannot read the address text, so it is not suitable for this logistics requirement.

Face API

Why wrong: Face API is a purpose-built service for analyzing human faces: it detects face landmarks, estimates age, emotion, and other attributes, and supports facial recognition tasks. It has no optical character recognition or text extraction capability, so it cannot read or interpret any characters, whether handwritten or printed. Since the task requires reading addresses—not detecting faces—Face API is entirely unrelated and would not produce the needed output.

Custom Vision

Why wrong: Custom Vision is a model training service that lets you create custom image classifiers or object detectors by uploading labeled images and training a model to recognize specific classes of objects. Its primary purpose is to detect predefined categories (e.g., 'defective part' or 'cat'), not to extract arbitrary textual content from images. To read handwriting, you would need to train a model for each character and still handle variability poorly, whereas Azure provides the Read API for OCR, making Custom Vision an inefficient and unsuitable choice here.