Study Azure AI-900 Responsible AI: key concepts, common traps, and exam decision cues.
Responsible-AI questions on AI-900 are usually short, but they are not filler. Microsoft wants you to identify which principle is most directly at risk in a scenario. That means you need clearer boundaries than “this feels unsafe.”
Accountability: Human responsibility for what the AI system does, including ownership, escalation, and oversight.
| Principle | What AI-900 wants you to notice |
|---|---|
| fairness | outcomes are unjustly biased across groups |
| reliability and safety | the system may fail unpredictably or cause harm |
| privacy and security | sensitive data, prompts, identities, or outputs are not protected well enough |
| inclusiveness | the system works poorly for certain populations, abilities, accents, or languages |
| transparency | users cannot tell AI is being used or cannot understand its limits |
| accountability | there is no human owner, review step, or escalation path for risky decisions |
| If the scenario is about… | Strongest first principle |
|---|---|
| certain applicants being scored lower for reasons unrelated to job fit | fairness |
| an AI assistant giving dangerous medical or safety advice | reliability and safety |
| sending sensitive customer records into a system without proper controls | privacy and security |
| speech recognition failing for certain accents or accessibility needs | inclusiveness |
| users not knowing an answer came from AI | transparency |
| no person being responsible for correcting harmful outputs | accountability |
Human review often signals accountability, because someone remains responsible for outcomes. But it can also support reliability and safety when the review exists to catch harmful failure modes before an AI output reaches a user. Read the reason for the review step, not just the presence of one.
| Problem | Strongest first mitigation at fundamentals depth |
|---|---|
| unfair outcomes | review data, evaluation patterns, and group-level outcome differences |
| harmful or unstable output | add testing, limits, safety controls, and fallback review |
| weak data protection | reduce exposure, secure access, and control sensitive inputs or outputs |
| poor performance for some users | test across varied users, languages, accents, and conditions |
| opaque AI use | disclose AI involvement and communicate limits clearly |
| no clear owner | define human ownership and escalation paths |