AI-900 Exam Questions and Answers

Fairness

Fairness is the AI principle that requires the model to treat all individuals and groups equitably by actively identifying and mitigating algorithmic bias, especially regarding protected attributes such as race, gender, or age. In a bank's loan approval system, fairness ensures that approval decisions do not disproportionately reject applicants from any demographic group, directly addressing the risk of discrimination. This is the correct focus because fairness specifically targets the elimination of bias in decision-making outcomes.

Reliability & Safety

Reliability & Safety is the Responsible AI principle concerned with ensuring an AI system performs consistently and without causing harm under real-world operating conditions. In a manufacturing context, predictive models must maintain high accuracy and robust behavior despite variations in environmental factors such as temperature, humidity, vibration, or sensor noise, while also guaranteeing that any automated decisions or alerts do not jeopardize worker or equipment safety. This principle directly matches the requirement for dependable, safe operation across different factory floors, rather than merely addressing data protection or social equity.

Model Fairness Assessment

Model Fairness Assessment directly evaluates the trained model's predictions across user-defined sensitive groups such as race, ethnicity, or gender. It computes fairness metrics like demographic parity, equalized odds, and disparate impact to quantify whether the model treats these groups unequally. Unlike data-focused checks, this component operates on model outputs, making it the appropriate tool for ensuring the resultant credit risk model does not encode discriminatory behavior.

Accountability

Accountability is a foundational principle of responsible AI, requiring that human beings remain responsible for the design, deployment, and outcomes of AI systems, especially in high-stakes domains like healthcare. Fully automating diagnosis removes the possibility of a human clinician to verify, override, or accept the AI's recommendation, thereby eliminating a clear line of responsibility for patient outcomes. Even if the system is technically reliable, accountability demands that there is a human who can be held responsible for the system's decisions and its consequences. This is the core issue in the scenario.

Transparency

Transparency is the correct principle because it directly addresses the customer's request for a meaningful explanation of why a loan or investment recommendation was made. In practice, transparency requires the AI system to provide interpretable justifications that trace how input data led to the output, often through techniques like feature importance or simplified decision paths. This supports user trust and satisfies compliance requirements such as GDPR's right to explanation, which explicitly obliges organizations to make automated decisions understandable.

Privacy and security

Privacy and security is the correct principle because it directly addresses how AI systems handle sensitive healthcare data. It requires implementing safeguards like encryption, access controls, and data minimization to prevent unauthorized access, theft, or misuse of patient records. Meeting standards such as HIPAA or GDPR depends on these measures, which align perfectly with protecting patient data in a clinical setting.

Regression

Regression predicts a continuous numeric value, which is exactly what is needed for predicting house price.

B. Recall

Recall, also called sensitivity or true positive rate, directly measures the proportion of actual positive instances that the model successfully identifies. It is computed as true positives divided by the sum of true positives and false negatives, i.e., 90 out of 90+10, which equals 0.9. This precisely matches the question's definition of 'the proportion of actual spam correctly identified.'

Principal Component Analysis (PCA)

Principal Component Analysis (PCA) is a dimensionality-reduction technique that uses orthogonal transformation to convert a set of possibly correlated features into a smaller set of linearly uncorrelated variables called principal components. These components are ordered so that the first few retain most of the variation present in the original data, allowing a high-dimensional feature set to be summarized with minimal information loss. By projecting customers onto a lower-dimensional subspace, PCA reduces the feature count while preserving the structure needed for predicting customer churn or purchase propensity.

Unsupervised learning - Clustering

Clustering is an unsupervised learning technique that partitions data points into groups based on feature similarity, such as purchase history, spending amount, or browsing behavior. Algorithms like k-means or DBSCAN automatically discover natural customer segments from the patterns in these features without requiring any labeled examples or predefined outcomes. This directly matches the retail company's goal of grouping customers based on their attributes.

Supervised classification

Supervised classification is appropriate because each patient record includes a known outcome label: either 'readmitted' or 'not readmitted' within a specified timeframe. The algorithm learns a decision boundary from historical features (demographics, diagnoses, lab values) to predict the categorical target for new patients. Common algorithms include logistic regression, random forests, and gradient-boosted trees, and the model is evaluated with metrics such as accuracy, precision, recall, and AUC-ROC.

Overfitting

The classic signature of overfitting is a model that achieves very high accuracy on training data but shows a substantial drop on validation or test data. Because the model has enough capacity to memorize the training examples—including their noise and outliers—it fails to generalize to unseen patterns. The large gap between training and validation performance is the defining diagnostic for overfitting.

Image classification

Image classification is the correct choice because it maps an entire input image to a single semantic label, such as 'delivery truck' or 'sedan', by evaluating the dominant visual features. The model is trained on labeled vehicle photos and outputs a probability distribution over the possible vehicle types, letting the transportation company quickly determine whether an image shows a particular category. This aligns directly with the requirement to identify the vehicle type without needing to localize objects or annotate individual pixels.

A. Use Custom Vision to train a classification or object detection model with transfer learning

Custom Vision is purpose-built for custom visual inspection: it applies transfer learning from a deep convolutional neural network pre-trained on generic image datasets (e.g., ResNet variants) to your specific defect photos. You only need to provide labeled examples of surface defects and acceptable parts; the service fine-tunes the model to classify each image or localize defects with bounding boxes via object detection. This is exactly the workflow required for detecting scratches, dents, or discoloration on metal surfaces, and it can be done with a relatively small dataset.

Optical character recognition (OCR)

Optical character recognition (OCR) directly extracts the textual characters from a shipping label image, converting handwritten names and addresses into machine-readable text. Azure AI Vision's Read API and OCR engines detect character shapes stroke-by-stroke, even for cursive or messy handwriting, and output the recognized words in a structured format. This is precisely what the logistics company needs to digitize thousands of handwritten labels.

Optical Character Recognition (OCR) using the Read API

The Azure AI Vision Read API performs OCR by detecting and extracting text from images, converting handwritten or printed characters into machine-readable strings. In this warehouse conveyor scenario, each serial number on a box label can be captured by a camera and transcribed into an alphanumeric value exactly as printed, even under uneven lighting, slight rotation, or varied label fonts. Because the goal is to record a specific identifier rather than categorize or describe the box, OCR is the only service that directly returns the serial number itself.

Azure Face API

Azure Face API is the correct choice because it provides a dedicated face verification endpoint that performs 1:1 matching. The service detects facial landmarks in a live photo, extracts a unique face embedding, and compares it against a reference photo (such as an ID document) to compute a similarity score against a configurable confidence threshold. This is precisely the identity-verification scenario the retail company needs.

Object detection

Object detection is a computer vision technique that identifies instances of defined classes within an image and localizes each instance with a bounding box. For a warehouse camera that must 'locate and frame' each box, object detection outputs the coordinates and dimensions of every detected box, enabling downstream tasks like robotic picking or inventory tracking. Unlike classification, it provides spatial localization rather than a single image-level label.

Text Analytics

Azure Text Analytics, part of the Cognitive Service for Language, includes Named Entity Recognition (NER) that automatically identifies predefined categories such as Person, Date, Quantity, and Organization in unstructured text. Additionally, the Text Analytics for Health feature extends this to clinical entities like symptoms, medications, and diagnoses, which is ideal for extracting specific data from medical notes. This prebuilt capability works on batches of documents without custom training, making it the correct choice.

Speech to Text and Text Analytics for Health

Speech to Text provides high-quality real-time transcription of doctor-patient conversations, converting spoken language into text. The Text Analytics for Health service is then applied to that transcript, using healthcare-specific NLP models to extract entities such as diagnoses, medications, symptoms, and treatment plans. Together, they form a purpose-built pipeline for clinical documentation, which is exactly what the hospital needs.

Language Understanding (LUIS)

LUIS, part of the Azure Language service, is the correct choice because it is expressly designed for natural language understanding in conversational AI. It performs intent recognition and entity extraction on user utterances, enabling a bot to map what a user says to a specific action (intent) and pull out key details (entities). LUIS also offers pre-built models for common intents, which accelerates bot development.

Key Phrase Extraction

Key phrase extraction is an Azure AI Language feature that returns a ranked list of the most salient words or multi-word expressions in a document. It uses statistical and semantic modeling to surface phrases like 'renewable energy adoption' or 'monetary policy,' which directly represent the important topics covered in an article. Because it is designed to summarize the main ideas rather than only detect tone, entities, or language, it is the correct service for automatically identifying what each daily article is about.

Custom Question Answering

Custom Question Answering (formerly QnA Maker) is a purpose-built Azure AI Language feature that ingests HR policy documents, PDFs, and FAQ pages to create a knowledge base. It uses semantic ranking and confidence scoring to match a user's natural-language query to the most relevant passage, then returns that passage as an answer. This directly satisfies the self-service bot requirement because the service is explicitly designed to answer questions from provided content.

Key phrase extraction

Key phrase extraction returns the main topics, themes, and descriptive phrases from text by analyzing linguistic features like term frequency and co-occurrence. In the Azure AI Language service, this API scans unstructured text and surfaces salient phrases such as 'long battery life' or 'camera clarity', making it the ideal choice for discovering what customers repeatedly mention about products.

A. Azure OpenAI Service

Azure OpenAI Service is the only option here that provides a managed, pre-built generative language model. It exposes APIs for GPT-4 and other autoregressive transformers that predict the next token in a sequence, allowing the marketing team to turn product keywords into coherent, human-readable descriptions with simple prompt instructions. These models support customization, content filters, and prompt engineering, so they can generate fresh copy immediately without any model training.

A. GPT-4 for text and DALL-E for images

GPT-4 is a large language model trained for multi-turn dialogue, enabling the bot to understand conversational context and generate coherent, human-like answers. DALL-E is a generative diffusion model that synthesizes original images from natural-language prompts, which is exactly what a chatbot needs to produce relevant visuals on demand. Together they provide both text generation and image generation, covering the full requirement.

Content filters

Azure OpenAI Service includes a set of configurable content filters that classify both prompts and completions into severity levels for categories such as hate, self-harm, sexual, and violent content. When a severity threshold is hit, the service blocks the response or returns a generic 'content filtered' message. This runtime safety control is the correct mechanism to prevent the model from generating offensive or inappropriate outputs, because it evaluates the actual semantic content in real time.

Configuring the content filtering (responsible AI filters)

Azure OpenAI’s content filtering system is a built-in safeguard that automatically screens inputs and outputs for categories like hate, violence, sexual content, and self-harm. Companies can configure severity levels to prevent undesirable content from being generated.

Retrieval Augmented Generation (RAG) — provide relevant document excerpts as context in the prompt

Retrieval Augmented Generation (RAG) is correct because it injects relevant document excerpts directly into the model's prompt at query time, grounding the response in the retrieved content rather than relying solely on the model's parametric memory. This context-supplemented prompting lets the model answer based on the provided evidence, significantly reducing hallucinations and allowing the response to be traced to specific source passages without any model retraining.

Temperature

Temperature directly scales the logits (raw scores) of every candidate token before the softmax layer is applied. A higher temperature flattens the probability distribution, allowing lower-ranked tokens to be selected more often, which results in more creative, varied, and even unpredictable text. Lowering temperature concentrates probability mass on the top tokens, making outputs more deterministic and coherent. This is the primary sampling setting Azure OpenAI exposes for controlling overall randomness vs. conservatism.