Use GitHub Copilot responsibly

Copilot can hand you a function that calls a library method that does not exist, written as confidently as a correct one. GitHub calls this a hallucination: output that sounds plausible but is factually incorrect, unsupported by the provided context, or entirely fabricated. This domain rests on one model: every Copilot output, whether a suggestion, a chat answer, a pull request or a commit message, is a proposal, and the person who accepts it answers for the result. The exam's recurring trap, covered below under "When two answers both work, pick the one that keeps a person checking", is an answer that lets something other than a person, such as a fluent explanation or a safety filter for harmful or insecure output (not the public code filter, whose Suggestions matching public code policy withholds matches on Block and shows them with a code reference on Allow), vouch for the output.

The first step, Understand Responsible AI Principles, explains why output is only a proposal. Copilot's limits trace back to three sources (this guide's grouping, not GitHub terminology): what the model learned, what it can see and how it generates. The six responsible AI principles (Fairness, Reliability and safety, Privacy and security, Inclusiveness, Transparency, and Accountability) name what a failure can harm. Reach for it when a scenario asks why output fails or which principle is at stake. The second step, Validate and Operate AI Tools, covers operation at three points (again this guide's grouping): before Copilot acts (while you are still writing the request), you scope it with context, a clear prompt and, for an agent task, the conditions a result must meet (acceptance criteria); while an agent acts (until its task finishes), you supervise it, including each action Copilot CLI asks you to approve; after Copilot acts, you validate its output, including a cloud agent's pull request. Reach for it when a scenario asks what to do next with a suggestion, command or pull request; the decision tree below routes each scenario.

Wrong answers in this domain usually pair a reasonable control with the claim that it is enough on its own. A check vouches only for the question it asks: a successful compile shows the code is syntactically valid, passing tests cover only the inputs they exercise, and a clean security scan finds only the kinds of problems it looks for (Validate and Operate AI Tools walks through each one). The same holds over time, because the last of the four functions in Microsoft's govern, map, measure, manage lifecycle, Manage, keeps monitoring and adjusting mitigations after release, so a one-time review before release is incomplete. The answer the exam rewards adds independent evidence and keeps an accountable person in the loop, with review depth scaled to what a missed defect would cost.