6.2 Practice Questions: AI Workloads and Responsible AI
Key Takeaways
- Domain 1 (15-20%) tests two skills: identifying the AI workload from a scenario and matching a scenario to a responsible AI principle.
- Match workloads by output: numbers = regression, known categories = classification, rare events = anomaly detection, images = computer vision, language = NLP, dialogue = conversational AI, new content = generative AI.
- The six responsible AI principles are fairness, reliability/safety, privacy/security, inclusiveness, transparency, and accountability — each with a distinct trigger.
- Responsible AI applies to ALL AI, not just generative AI — fairness in ML models, privacy in NLP, inclusiveness in speech.
- Azure AI Content Safety is the concrete harmful-content service; the six principles are the broader governance framework.
Test your knowledge of Domain 1 with these scenario-based questions. Each mirrors the style and difficulty of the actual AI-900 exam, where Domain 1 (AI workloads and considerations) is 15-20% of the questions.
What Domain 1 Tests
Domain 1 has two halves, and most questions fall into one of them:
- Identify the AI workload from a scenario.
- Identify the responsible AI principle that a scenario supports or violates.
For workload questions, decide what the system produces from its input:
| Clue in the scenario | Workload |
|---|---|
| Predicts a number (price, demand, temperature) | Regression / prediction |
| Sorts items into known categories (spam/not spam, defect/no defect) | Classification |
| Flags rare or unexpected events in a stream | Anomaly detection |
| Interprets images or video (objects, faces, text in images) | Computer vision |
| Understands or generates written/spoken language | NLP |
| Holds a multi-turn dialogue with users | Conversational AI |
| Creates new content (text, images, code) | Generative AI |
Database administration and network configuration are NOT AI workloads — they are classic distractors.
The Six Responsible AI Principles
Responsible AI is the conceptual backbone of Domain 1, and it reappears in every other domain. Memorize the six principles and a one-line trigger for each:
| Principle | Trigger phrase in a question |
|---|---|
| Fairness | Discriminates against a group; biased outcomes; proxy variables (zip code standing in for race) |
| Reliability and Safety | Must work consistently and safely; rigorous testing; safe failure (self-driving, medical devices) |
| Privacy and Security | Personal data collected/stored; encryption and access control; consent |
| Inclusiveness | Must work for people of all abilities and backgrounds; accessibility |
| Transparency | Users must understand how/why a decision was made; explainability |
| Accountability | A human or organization is answerable; governance; oversight |
A useful tie-breaker: fairness is about who is treated unequally, transparency is about why a decision cannot be explained, reliability/safety is about whether it works correctly and safely, and privacy/security is about protecting data. When two principles seem to fit, pick the one the scenario emphasizes most directly.
Also keep Azure AI Content Safety distinct from the principles themselves: Content Safety is the concrete service that detects harmful content (hate, violence, sexual, self-harm) across text and images, whereas the six principles are the broader governance framework.
Worked Reasoning: Two Hard Cases
Case 1 — fairness vs. transparency. A hiring model rejects more candidates from one demographic. Is this fairness or transparency? The scenario emphasizes unequal outcomes for a group, so the answer is fairness. It would only be transparency if the complaint were that no one could explain why a candidate was rejected. Read for the emphasis: outcomes -> fairness; explainability -> transparency.
Case 2 — anomaly detection vs. classification. A system flags fraudulent credit-card transactions. If the system sorts each transaction into a known label (fraud / not fraud) from labeled history, it is classification. If it flags anything that deviates from a customer's normal behavior without predefined fraud labels, it is anomaly detection. The presence or absence of predefined labels decides it — the same logic that separates classification from clustering in Domain 2.
Mapping Workloads to Azure Services
Domain 1 stays conceptual, but it helps to connect each workload to the Azure service you will see later:
| Workload | Typical Azure service |
|---|---|
| Prediction / regression / classification | Azure Machine Learning |
| Anomaly detection | Azure Machine Learning (and Metrics Advisor patterns) |
| Computer vision | Azure AI Vision / Custom Vision |
| NLP | Azure AI Language |
| Conversational AI | Azure AI Language (CLU) + Bot Service |
| Generative AI | Azure OpenAI Service |
| Harmful-content detection | Azure AI Content Safety |
How to Use This Practice Set
Treat these questions as an active-recall checkpoint, not a reading assignment. Answer the full set before looking at explanations, then mark each miss by skill area, rule, or service name. For every wrong answer, write down why the correct option wins and why one tempting distractor fails. That habit matters because the real exam tests the same concept with different wording. If you miss several questions from the same domain, pause and reread that chapter before continuing.
Review Routine
After you finish, build a three-column log: topic, the rule you missed, and the clue you should have noticed. Retake only missed questions the next day, then mix them with new questions so you do not simply memorize order. For scenario questions, name the workload or principle aloud before choosing an option; that prevents rushing into a familiar but wrong answer. A strong final loop is: timed attempt, explanation review, targeted reread, then a second attempt after a short break.
Finally, remember that Domain 1 questions are deliberately scenario-driven: they describe a real situation and ask you to abstract it to a workload or a principle. Practicing that abstraction — stripping a story down to "what does it produce?" or "who or what is harmed, and how?" — is the single most transferable skill for the entire exam, because the other four domains lean on the same habit.
A bank uses an AI model to approve or deny loan applications. The model consistently denies applications from a specific zip code that correlates with a minority neighborhood, even when applicants have strong credit histories. Which responsible AI principle is being violated?
A retail company wants to predict how many units of each product they will sell next month to optimize inventory. Which type of AI workload is this?
An autonomous vehicle uses AI to identify and locate pedestrians, cars, traffic signs, and lane markings in real time. Which two computer vision tasks are primarily involved?
A medical AI system provides a diagnosis but cannot explain how it reached its conclusion. The doctor cannot determine why the AI recommended a specific treatment. Which responsible AI principle is lacking?
A cybersecurity system monitors network traffic and flags unusual patterns that might indicate a breach. This is an example of which AI workload?
Which statement about AI and machine learning is correct?
An AI customer service chatbot collects and stores customer conversations including personal information, but the company has no data protection policy or encryption for this data. Which responsible AI principle is most at risk?
Which THREE are common types of AI workloads? (Select three)
Select all that apply