Microsoft Azure AI Fundamentals AI-900 (AI-900) — Questions 1–75

Key Phrase Extraction identifies important terms and concepts in text, but it does not detect or redact PII like names, addresses, or social security numbers. The question requires a feature that specifically identifies PII for redaction.

Sentiment Analysis determines the emotional tone of text (positive, negative, neutral), not the presence of PII. The law firm needs to identify and redact specific data like names and SSNs, which Sentiment Analysis cannot do.

Language Detection identifies the language of text, not PII. The requirement is to redact PII, not determine the language of the documents.

A company needs to automatically extract the main topics or keywords from customer feedback surveys to identify common themes. They want a prebuilt Azure AI Language feature without custom training. Key Phrase Extraction would be the correct choice.

A company wants to automatically gauge customer satisfaction from support tickets by analyzing the emotional tone of the text. Sentiment Analysis would be the correct prebuilt feature to classify each ticket as positive, negative, or neutral.

A multinational company needs to automatically route customer support tickets to language-specific teams based on the language of the message. Language Detection would be the correct choice to identify the language of each ticket.

Candidates may confuse 'key phrases' with 'personal information,' thinking that extracting important terms includes names and addresses, not realizing PII detection is a separate, specialized feature.

Candidates may confuse 'analyzing text for sensitive content' with 'analyzing text for emotion,' both involving text analytics, but they serve different purposes.

Candidates may confuse Language Detection with PII Detection because both are prebuilt Azure AI Language features, and they might think detecting language is a prerequisite for processing documents, but the question explicitly asks for PII redaction.

The kiosk uses only a voice interface, excluding people who are deaf or hard of hearing, which violates the Inclusiveness principle, not Accountability.

The kiosk's voice-only interface excludes users with hearing impairments or those who rely on text/visual interfaces, directly violating the accessibility principle. Option B does not address this core issue.

The kiosk's lack of text or screen reader support directly violates the accessibility principle, not privacy and security. Privacy and security are about data protection and system safety, which are not the primary issues here.

A company deploys an AI system that makes loan approval decisions. A customer asks for an explanation of why their loan was denied, but the company cannot provide any reasoning. Which principle is violated?

Option B would be correct in a scenario where an AI system produces biased outcomes against a specific demographic group, such as a hiring algorithm that discriminates based on gender or race, directly violating the fairness principle.

This option would be correct in a scenario where an AI system collects sensitive personal data without encryption or consent, or where there is a data breach due to inadequate security measures.

Candidates may confuse the need for accountability in AI systems with the accessibility issue, thinking that the lack of alternative input methods is a failure of accountability rather than inclusiveness.

Candidates may confuse accessibility with fairness, thinking that excluding some users is a fairness issue rather than an accessibility one, or they may not clearly distinguish between the two principles.

Candidates may confuse accessibility with privacy/security because both involve protecting users, but they focus on different aspects: one on inclusive design, the other on data protection.

Temperature controls randomness of output, not content safety; it cannot filter profanity, hate speech, or self-harm references.

Top-p parameter controls nucleus sampling for output diversity, not content safety; it cannot filter profanity, hate speech, or self-harm references.

Max-tokens parameter controls the length of generated text, not content safety. It cannot block harmful content like profanity or hate speech.

When the question asks how to control the creativity or variability of generated text (e.g., 'Which parameter should be adjusted to make captions more deterministic?').

A question asking how to control the randomness and diversity of generated text, such as 'Which parameter should be adjusted to make the model's responses more focused and less random?' would make Top-p the correct answer.

A question asking how to limit the number of tokens in a model's response to reduce cost or ensure concise output would have max-tokens as the correct answer.

Candidates may confuse parameters that influence output style with safety controls, assuming 'temperature' affects content appropriateness.

Candidates may confuse parameters that influence output quality with safety mechanisms, assuming that adjusting sampling parameters can also block harmful content.

Candidates may mistakenly think that limiting output length (max-tokens) can prevent harmful content, confusing quantity control with content filtering.

Image Analysis provides general information about images (objects, colors, etc.) but does not extract specific text like addresses or names from handwritten forms.

Face detection identifies human faces in images, but the question requires extracting text (address, name, weight) from handwritten forms, which is not a face-related task.

Custom Vision is used for training custom image classification or object detection models, not for extracting text from images. The question specifically requires extracting handwritten text, which is a text extraction task, not a custom vision model task.

If the question asked for identifying objects or describing the content of an image (e.g., 'Which capability can detect that a photo contains a dog and a ball?'), Image Analysis would be correct.

A question asking for a solution to detect and locate faces in images for security or demographic analysis, such as counting people in a crowd or verifying identity in a photo.

A company needs to identify specific objects (e.g., logos, package types) in shipping forms, and the available pre-built models do not cover those objects. Custom Vision would allow training a custom object detection model to identify those specific items.

Candidates may confuse Image Analysis with OCR because both process images, but Image Analysis lacks the text extraction capability needed for handwritten forms.

Candidates may confuse 'detection' with 'extraction' or think face detection can read text, or they might assume any form of detection is suitable for processing forms.

Candidates may think 'Custom Vision' can be trained to read handwriting, but OCR is the dedicated service for text extraction; Custom Vision focuses on visual features, not text recognition.

Fairness is about ensuring AI systems treat all people equitably and avoid bias, but the question specifically asks about the right to understand how a decision was made, which is a transparency concern.

The question focuses on the right to understand how a model arrived at its decision, which directly relates to transparency. Reliability and safety concern system performance and robustness, not explainability.

The question focuses on students' right to understand how the model arrived at its decision, which directly relates to transparency, not privacy and security. Privacy and security would be relevant if the concern were about unauthorized access or misuse of student data.

A university deploys an AI model to predict student dropout risk, and it is discovered that the model systematically flags students from a particular demographic group more often than others. Which principle is most directly relevant?

A hospital uses an AI system to diagnose diseases. The system must consistently produce accurate results and fail safely. Which principle ensures the system performs reliably under all conditions?

A healthcare organization uses an AI system to predict patient readmission risks, and patients are concerned about their medical data being exposed or used without consent. The most relevant principle would be privacy and security.

Candidates may confuse fairness with transparency because both involve ethical AI, and they might think that understanding the model's decision is related to ensuring it is fair.

Candidates may confuse the need for the model to be 'reliable' in its predictions with the need for the model's decisions to be 'explainable,' or they may think that understanding the decision is part of ensuring reliability.

Candidates may confuse the right to understand a model's decision (transparency) with data protection (privacy), or they may assume that any student-related AI system inherently involves privacy concerns.

Optical Character Recognition (OCR) extracts text from images, not objects like people or pets. The requirement is to detect objects and their bounding boxes, not text.

Image Captioning generates a textual description of the entire image, but does not identify specific objects with bounding boxes, which is required for detecting presence and location of objects like a person or package.

Custom Vision requires custom training with labeled images, but the question specifies using a prebuilt Azure AI service without any custom training.

A question asks: 'A system needs to extract printed text from scanned documents for digital processing. Which Azure Computer Vision capability should be used?' OCR would be correct.

A question asking for generating a natural language description of an image's content, such as 'Describe what is happening in this photo' for accessibility or content moderation, where no object localization is needed.

A scenario where the home security system needs to detect unique objects not covered by prebuilt models, such as specific brand packages or rare pet breeds, and the team is willing to provide labeled training images.

Candidates may confuse OCR with general image analysis or think that detecting text on packages (e.g., labels) is the primary need, overlooking the object detection requirement.

Candidates may confuse 'detecting objects' with 'describing objects' and assume captioning can provide object presence, not realizing it lacks spatial localization (bounding boxes).

Candidates may think Custom Vision is the only way to detect objects, not realizing that prebuilt object detection in Image Analysis already covers common household objects like people, pets, bags, and packages.

Temperature controls randomness of token selection, not the probability of specific tokens. It cannot prevent a particular word like 'Corporation' from being generated.

Top P (nucleus sampling) controls the cumulative probability threshold for token selection, not the probability of a specific token. It cannot be used to reduce the likelihood of a particular word like 'Corporation'.

Frequency penalty reduces the likelihood of tokens based on how often they have appeared in the generated text so far, but it does not allow targeting a specific word like 'Corporation' for exclusion. The question requires preventing a specific token regardless of its frequency, which is achieved by logit bias.

A question asking how to make model outputs more deterministic (e.g., 'Which parameter should be set to 0 to always choose the most likely token?') would have Temperature as the correct answer.

A developer wants to generate diverse product names but avoid overly rare or nonsensical suggestions. They set Top P to 0.9 to sample only from tokens that make up the top 90% of probability mass, ensuring outputs are coherent while maintaining variety.

A developer wants to reduce repetitive language in a long-form text generation task, such as a story or article, where the model keeps repeating the same phrases. Frequency penalty would be the correct parameter to decrease the probability of tokens that have already been generated frequently.

Candidates may think temperature affects all tokens equally and can be tuned to avoid certain words, confusing overall randomness with token-specific suppression.

Candidates may confuse Top P with a mechanism to filter out specific tokens, as it does restrict the set of possible tokens, but it does so based on cumulative probability, not individual token identities.

Candidates may confuse frequency penalty with a mechanism to suppress specific words, as both involve reducing token probabilities. However, frequency penalty is based on occurrence count in the output, not on a predefined list of forbidden tokens.

The question focuses on system failure due to untrained weather conditions, which directly impacts reliability and safety, not fairness. Fairness relates to bias against demographic groups, not environmental conditions.

The question describes a failure in system performance under novel conditions (snow/fog), which directly violates the reliability and safety principle. Transparency concerns explainability or disclosure of system behavior, not performance degradation.

The question focuses on system failure due to environmental conditions not seen in training data, which directly relates to reliability and safety, not to privacy or security. Privacy and security concerns involve data protection and unauthorized access, not performance in adverse weather.

Fairness would be correct if the system performed poorly for certain demographic groups (e.g., pedestrians with darker skin tones) due to biased training data, leading to unequal safety outcomes.

Transparency would be the correct answer if the question described a system that operates correctly but fails to provide explanations for its decisions, or if the company did not disclose the system's limitations (e.g., not informing users that the system was trained only on sunny data).

This option would be correct in a scenario where an AI system exposes sensitive user data (e.g., facial recognition system leaking biometric information) or is vulnerable to adversarial attacks that compromise data confidentiality.

Candidates may confuse 'unfair' outcomes (unequal performance across conditions) with the fairness principle, but fairness in AI refers to demographic parity, not environmental robustness.

Candidates may confuse the lack of disclosure about training data limitations (a transparency issue) with the actual safety failure, or they may think that explaining why the system fails would have prevented the problem.

Candidates may confuse 'safety risks' with 'security risks' or think that deploying in new conditions creates a privacy issue, but the core problem is system reliability, not data protection.

Key phrase extraction identifies general phrases (e.g., 'contract date') but does not classify them into predefined categories like dates, party names, or monetary amounts. The question requires extracting specific entity types, which is the function of NER.

Sentiment analysis determines the emotional tone (positive, negative, neutral) of text, not the extraction of specific entities like dates, names, or monetary amounts.

Language detection identifies the language of text (e.g., English, Spanish), not specific terms like contract dates or party names. The firm needs to extract predefined information, not detect the document's language.

A company wants to automatically extract the most important topics or keywords from customer feedback without needing to classify them into predefined categories. Key phrase extraction would be the correct choice.

A company wants to automatically gauge customer satisfaction from product reviews by identifying whether the feedback is positive, negative, or neutral. Sentiment analysis would be the correct prebuilt feature.

A multinational company receives documents in multiple languages and needs to automatically route them to the appropriate language-specific processing pipeline. Language detection would be the correct feature to identify each document's language.

Candidates may confuse 'extracting terms' with 'extracting key phrases,' not realizing that NER is specifically designed to identify and classify predefined entity types like dates and names.

Candidates may confuse 'extracting terms' with 'analyzing text' and think sentiment analysis can identify specific information, not realizing it only assesses overall sentiment.

Candidates may confuse 'language' with 'information extraction' or think that language detection is a prerequisite step before applying other features, but the question specifically asks for extracting terms, not identifying language.

The disparity in triage priority based on symptom rarity is a fairness issue, not an inclusiveness issue. Inclusiveness focuses on ensuring diverse user groups can access and use the system, not on equitable treatment outcomes.

The disparity in triage priority is an issue of bias, not lack of transparency. Transparency concerns understanding how the system works, not ensuring equitable treatment.

Accountability focuses on assigning responsibility for AI system outcomes, not on the specific issue of bias or unequal treatment. The disparity described is a fairness problem, not a lack of accountability.

A question where the AI system is designed for a diverse population but fails to accommodate users with disabilities or language barriers, such as a health chatbot that only supports English and lacks screen reader compatibility. In that case, inclusiveness would be the most relevant principle.

A healthcare clinic uses an AI system to triage patients, but the system's decision-making process is a black box. The clinic wants to understand why certain patients receive higher priority. In this scenario, Transparency would be the most relevant principle.

A healthcare clinic deploys an AI diagnostic tool that makes errors, and patients are harmed. The clinic wants to ensure there is a clear process for identifying who is responsible for the system's decisions and for remedying harm. In that scenario, Accountability is the most relevant principle.

Candidates may confuse 'inclusiveness' with 'fairness' because both relate to equitable treatment, but inclusiveness specifically addresses accessibility and representation of diverse groups, not bias in decision outcomes.

Candidates may confuse the need to explain the system's behavior (transparency) with the need to correct its biased outcomes (fairness).

Candidates may confuse the need for the clinic to 'take responsibility' for the biased outcomes with the principle of Accountability, not realizing that the root cause is a fairness violation, not a lack of oversight or blame assignment.

The pre-built Computer Vision object detection API is not trained on penguins in rocky terrain and would fail to accurately detect partially visible, small penguins blending into the background, especially with only 500 labeled images for fine-tuning.

Image classification assigns a single label to the entire image, not detecting or localizing individual objects. Since the team needs to count penguins that are small and partially visible, object detection is required to identify each penguin's location.

A question where the objects are common (e.g., cars, dogs) and the user has no labeled data, requiring a zero-shot solution. For example: 'A company needs to detect common office items in photos without any labeled training data.'

If the team only needed to determine whether penguins are present in an image (e.g., presence/absence monitoring) and did not require counting or localization, training a Custom Vision image classification model would be appropriate.

Candidates may assume pre-built APIs work for any object detection task without considering the need for domain-specific training on rare or camouflaged objects.

Candidates may confuse image classification with object detection, assuming that classification can also count objects, or they may not fully understand the difference between the two tasks.

OCR extracts text from images, not background removal. The question asks for removing backgrounds from product photos, which is a segmentation task, not text extraction.

Image Captioning generates descriptive text for an image, not background removal. The question specifically requires removing the background from product photos, which is a segmentation task, not a captioning task.

Object Detection identifies and locates objects in an image, but does not remove backgrounds. The question specifically requires background removal, which is a distinct capability.

An exam question like 'Which Azure Computer Vision capability should be used to extract printed text from scanned invoices?' would make OCR the correct answer, as it specializes in text recognition.

An exam question asking: 'Which Azure Computer Vision capability can automatically generate a human-readable description of a product photo for accessibility purposes?' would make Image Captioning the correct answer.

An exam question asking: 'Which Computer Vision capability should be used to count the number of products in a warehouse photo and draw bounding boxes around each product?' would make Object Detection correct.

Candidates may confuse OCR with general image processing, thinking it can 'clean up' images by removing text or other elements, but OCR only reads text, not removes backgrounds.

Candidates may confuse 'Image Captioning' with general image processing capabilities, assuming it can modify images, or they may think captioning involves editing the image itself rather than describing it.

Candidates may think object detection can isolate products from the background, confusing detection (locating objects) with segmentation (separating objects from background).

Temperature controls randomness of output, not the number of completions. Setting temperature does not determine how many separate responses are returned.

The 'max_tokens' parameter controls the maximum number of tokens (words/punctuation) in a single completion, not the number of completions returned. Setting it to 5 would limit each slogan to about 5 tokens, not return 5 slogans.

The 'stop' parameter specifies sequences where the API should stop generating further tokens, not the number of completions. It does not control how many separate text outputs are returned.

A question asks: 'Which parameter should be adjusted to make the generated text more creative and less deterministic?' In that case, temperature would be correct.

A developer wants to ensure each generated slogan is concise, no longer than 50 tokens. The question asks which parameter limits the length of each completion. In that scenario, 'max_tokens' is the correct answer.

A developer wants to generate a product slogan but needs the generation to stop when a specific phrase like 'END' appears, to avoid extra text. Setting the 'stop' parameter to ['END'] would be correct.

Candidates may confuse temperature with controlling output diversity, thinking higher temperature yields more distinct options, but it only affects randomness per completion.

Candidates may confuse 'max_tokens' with controlling the number of outputs, as both involve a numeric value. They might think setting it to 5 returns 5 results, misunderstanding that it limits length per result.

Candidates may confuse 'stop' with controlling when generation ends, mistakenly thinking it limits the number of completions rather than halting a single completion.

Increasing the frequency penalty reduces repetition of phrases but does not prevent the model from inventing facts; it does not ground responses in provided data.

Content filtering blocks harmful or biased language but does not prevent the model from inventing plausible-sounding incorrect details (hallucinations) unrelated to the provided data.

A developer wants to reduce repetitive language in AI-generated marketing copy while still allowing creative variation. Increasing the frequency penalty would discourage the model from reusing common phrases, making outputs more diverse.

A developer needs to ensure that generated responses for a customer-facing chatbot do not contain offensive, biased, or inappropriate content, even if the model otherwise produces accurate information.

Candidates may think that penalizing frequency will also penalize hallucinated content, but frequency penalty only affects token repetition, not factual accuracy.

Candidates may confuse content filtering with factuality controls, assuming that blocking certain outputs also prevents hallucinations, when in fact it only addresses safety and policy violations.

The core issue is unauthorized use of personal data (names, addresses, phone numbers) without consent, which directly violates privacy and security principles. Fairness is about bias and discrimination, not data protection.

The concern is about using personal data without consent, which violates privacy and security, not reliability and safety. Reliability and safety would be relevant if the system made errors or caused harm, but the issue here is data misuse.

Transparency is about informing users about AI system capabilities and limitations, but the core violation here is using personal data without consent, which directly breaches privacy and security principles.

A question where an AI system consistently gives lower credit scores to a specific ethnic group despite similar financial histories, or where a hiring algorithm rejects female candidates more often than males for the same qualifications.

A medical diagnosis AI system occasionally misdiagnoses rare conditions due to insufficient training data, leading to patient harm. In that scenario, reliability and safety would be the most directly violated principle.

A question where an AI system provides recommendations without explaining how they are generated, and customers are unaware of the data used. For example: 'A bank uses an AI to approve loans but does not disclose the factors influencing decisions.' Then transparency is the most violated principle.

Candidates may confuse 'unfair' data practices (using personal info without consent) with the Fairness principle, which addresses bias and equitable treatment, not data privacy violations.

Candidates may confuse 'safety' with data protection, or think that using personal data unsafely (e.g., data breach) falls under reliability and safety, but the core issue is unauthorized data collection and use.

Candidates may confuse the lack of customer notification about data use (which is a transparency issue) with the actual violation of unauthorized data collection and processing, which falls under privacy and security.

OCR extracts text from images, but the question requires detecting and counting items on a shelf without identifying product type, which is not text-based.

Semantic segmentation assigns a class label to every pixel in an image, which is overkill for simply detecting and counting items on a shelf; object detection is more efficient for identifying and localizing individual items without needing pixel-level precision.

A question where the goal is to read product labels, expiration dates, or barcodes from shelf items to track inventory by text.

A question that requires precise measurement of shelf space occupancy or identifying the exact shape and boundaries of items for inventory layout optimization would make semantic segmentation the correct choice, such as 'Which Computer Vision capability should be used to calculate the percentage of shelf space occupied by products?'

Candidates may confuse inventory monitoring with reading product labels, assuming OCR is needed to identify items.

Candidates may confuse semantic segmentation with object detection because both involve identifying objects, but segmentation's pixel-level detail seems more 'precise' for counting, even though it's unnecessary for simple item detection and counting.

Inclusiveness focuses on designing systems that are accessible and usable by people of all abilities and backgrounds, but it does not directly address the requirement to detect and mitigate statistical bias after removing protected attributes. The question specifically asks about investigating remaining bias, which falls under Fairness.

Reliability and safety focuses on system dependability and resilience to failures or adversarial attacks, not on addressing bias that persists despite removing protected attributes.

Transparency focuses on explainability and openness about system behavior, not on actively mitigating bias. The question asks which principle requires investigating and addressing bias, which is the core of Fairness.

Inclusiveness would be correct if the question asked: 'Which principle emphasizes designing AI systems that empower everyone, including people with disabilities, and ensuring diverse user needs are considered throughout the development process?'

An AI system for autonomous driving occasionally fails to detect pedestrians in low-light conditions, leading to safety risks. The question asks which principle requires rigorous testing and monitoring to ensure consistent performance under varied conditions.

Transparency would be correct if the question asked about the principle that requires the company to document and communicate how the AI system makes decisions, including disclosing limitations and potential biases to stakeholders.

Candidates may confuse inclusiveness with fairness because both relate to equitable treatment, but inclusiveness is broader and less focused on statistical bias detection and mitigation.

Candidates may confuse 'bias' with 'unreliable behavior' and think that addressing bias falls under ensuring the system works correctly for all groups, but reliability and safety is about operational robustness, not fairness.

Candidates may think that being transparent about bias (i.e., acknowledging it) fulfills the requirement, but Transparency does not mandate corrective action—only disclosure.

Sentiment Analysis detects positive, negative, or neutral sentiment in text, but it cannot ingest PDF documents or answer questions based on policy content.

Key Phrase Extraction identifies important terms in text but does not enable a question-answering system over documents; it lacks the ability to map questions to answers from ingested PDFs.

Language Detection identifies the language of text, but the question requires extracting answers from policy documents, not detecting their language.

A company wants to analyze customer reviews to determine overall satisfaction with a product. Sentiment Analysis would be the correct feature to classify each review's sentiment.

An exam question asks: 'A company wants to automatically extract the most important topics from thousands of customer reviews to identify common issues. Which Azure AI Language feature should they use?'

A company receives customer feedback in multiple languages and needs to route each message to the appropriate language-specific support team. Language Detection would be used to identify the language of each message.

Candidates may confuse sentiment analysis with understanding document content, thinking it can extract meaning from policy documents to answer questions.

Candidates may think extracting key phrases from policy documents is sufficient to answer questions, confusing information retrieval with question answering.

Candidates might think language detection is necessary to process multilingual documents, but the question focuses on answer extraction, not language identification.

Named Entity Recognition (NER) identifies and categorizes entities like people, organizations, and locations, but it does not extract multi-word phrases summarizing main topics such as 'product quality' or 'shipping delay' from unstructured text.

Language Detection identifies the language of text (e.g., English, Spanish), not topics or concepts within the text. The analyst needs to summarize main topics like 'product quality', not detect the language of the feedback.

Sentiment Analysis determines the emotional tone (positive, negative, neutral) of text, but does not extract or summarize topics or concepts like 'product quality' or 'shipping delay'.

A question asks: 'Which Azure AI Language feature should be used to extract specific entity types (e.g., person names, dates, monetary values) from legal documents?' In that case, NER is the correct answer because it is designed to identify and classify named entities.

A company receives customer feedback in multiple languages (e.g., English, French, German) and needs to route each message to the appropriate language-specific support team. Language Detection would be the correct feature to identify the language of each feedback email.

A question asking to determine whether customer feedback emails are generally positive, negative, or neutral about a product or service, without needing to identify specific topics.

Candidates may confuse 'entities' with 'topics' or 'key phrases', assuming that extracting named entities would also capture recurring concepts, but NER focuses on specific categories rather than summarizing main themes.

Candidates may confuse 'detecting topics' with 'detecting language', or think Language Detection can extract themes because it analyzes text content, but it only identifies the language.

Candidates may confuse sentiment analysis with topic extraction, thinking that identifying positive/negative sentiment also reveals the underlying topics, but sentiment analysis only provides polarity, not subject matter.

Azure Face API is designed for facial recognition and analysis, not for extracting text from documents, especially those with mixed printed and handwritten content.

Azure Custom Vision is designed for image classification and object detection, not for extracting text from documents, especially mixed printed and handwritten text.

Azure Video Indexer is designed to extract insights from video and audio content, not from scanned documents like invoices. It cannot process text from images or handwritten notes.

A question asking which Azure service can detect and identify individuals in a photo or video feed, such as for security or attendance tracking, would make Face API the correct answer.

A company needs to classify scanned invoices into categories (e.g., 'paid' vs 'unpaid') based on visual features like logos or stamps, without extracting text content.

A question asking which Azure service extracts metadata, transcripts, and spoken content from recorded meetings or video files would make Azure Video Indexer the correct answer.

Candidates may confuse 'extract text' with 'extract features' or mistakenly think Face API can handle any visual data extraction, including text from documents.

Candidates may think Custom Vision can handle any image-based task, including text extraction, due to its name implying broad visual recognition capabilities.

Candidates may confuse 'extracting information from media' with 'extracting text from documents,' assuming Video Indexer handles any unstructured data including scanned images.

The system's lack of explanation for loan decisions violates transparency, not fairness. Fairness would be violated if the model exhibited bias against protected groups, but the question focuses on the absence of reasoning, not on discriminatory outcomes.

The system's lack of explanation for loan decisions directly violates transparency, not reliability and safety. Reliability and safety concern system accuracy and robustness, which are not questioned here.

The question focuses on the lack of explanation for loan decisions, which directly violates transparency. Privacy and security are not the primary issue here, as no data breach or misuse of personal information is described.

This option would be correct if the question described an AI system that consistently rejects loan applications from a specific demographic group (e.g., based on race or gender) without legitimate justification, directly violating the fairness principle.

This option would be correct if the question described an AI system that frequently makes incorrect loan approvals or rejections due to model errors, or fails to handle edge cases safely, leading to financial harm.

This option would be correct if the question described a scenario where the AI system exposes sensitive applicant data without consent, or fails to protect loan application information from unauthorized access, violating privacy and security principles.

Candidates may associate loan approval systems with fairness concerns (e.g., bias against minorities) and overlook that the core issue here is the lack of explanation, which falls under transparency.

Candidates may confuse 'lack of explanation' with 'unreliable' because they assume an opaque system is more likely to be wrong, but the principle violated is transparency, not reliability.

Candidates may confuse the lack of explanation with a privacy concern, thinking that not providing reasons is a form of hiding information, which they incorrectly associate with privacy violations.

The target variable is the number of units sold, which is a continuous value, not a binary outcome. Binary classification predicts one of two discrete classes, not a continuous number.

The target variable is the number of units sold, which is a continuous value, not a discrete class. Multiclass classification predicts a single label from multiple discrete categories, not a continuous number.

Clustering is an unsupervised learning task used to group similar data points without labeled outcomes, but this question requires predicting a continuous target variable (units sold) from labeled historical data, which is a supervised regression problem.

A question where the goal is to predict whether sales will exceed a certain threshold (e.g., 'Will sales be above 1000 units?') with a yes/no outcome, using historical data and features like promotions and holidays.

A company wants to predict which of five product categories (e.g., electronics, clothing, food, books, toys) a new product belongs to based on its features. The target is a discrete category, making multiclass classification appropriate.

A company wants to segment its customers into distinct groups based on purchasing behavior, demographics, and browsing history, without predefined labels, to tailor marketing strategies. This is a clustering task.

Candidates may confuse predicting a specific number with a simple yes/no outcome, or they might think that any prediction involving sales is a classification problem.

Candidates may confuse 'multiple possible outcomes' (like different sales numbers) with 'multiple classes,' not realizing that regression handles continuous outcomes while classification handles discrete labels.

Candidates may confuse clustering with regression because both involve analyzing patterns in data, but clustering lacks a target variable and is unsupervised, while regression requires labeled continuous outcomes.