Using Generative AI at Work: Verifying, Catching and Handling Wrong Output

This behavioral round focuses on how you have actually used generative AI tools, such as code assistants and chat models, in your past work. The interviewer asks how you judge whether an AI-generated result is correct, how you have discovered that an AI result was wrong, and what you did once you found the problem, then follows up in depth on the real situations, your verification methods and how you handled them. ### Clarifying Questions - Does any use count (writing code, tests, documentation, debugging), or should the story be about code that shipped? - Should I describe one main situation in depth, or several shorter ones? - Are school or personal projects acceptable if my professional use has been limited? ### Part 1 — How you use generative AI and judge its output Describe how you use generative AI in your work and how you decide whether a result is correct before you rely on it. ```hint Name your checks Think about the concrete checks you run on generated code or answers, and how they change with the risk of the task and with how well you know the area. ``` #### What This Part Should Cover - Concrete, real use cases, and the places where you deliberately do not use AI - Specific verification methods matched to the kind of output - Scrutiny that scales with risk and with your familiarity with the area ### Part 2 — Discovering a wrong result Describe a specific time an AI tool gave you a wrong result. How did you find out? ```hint What gave it away? Recall the moment you noticed: a failing test, a reviewer's comment, behavior at runtime, or something that looked plausible but contradicted the documentation. ``` #### What This Part Should Cover - A specific incident and the actual nature of the error - How and how early it was detected - Why the error was plausible enough to get past a first look ### Part 3 — Handling the problem What did you do after you discovered the problem, both to fix it and to keep it from happening again? ```hint Beyond the fix Think about what changed in your workflow afterwards, and who else needed to know. ``` #### What This Part Should Cover - How the immediate issue was fixed and its impact contained - Taking ownership of AI-produced work as your own - Lasting changes to tests, review habits, prompts or team guidance ### What a Strong Answer Covers - Real, specific experience rather than general opinions about AI - A verification approach proportional to risk, with concrete techniques - Accountability for AI-generated output as if you had written it yourself - A learning loop: a process that changed after a mistake - An honest view of where AI helped and where it did not ### Follow-up Questions - How do you verify AI output in a part of the codebase or a language you do not know well? - When would you decide not to use an AI tool for a task at all? - How would you review a teammate's change that you suspect was mostly generated by AI? - What would you do if AI-generated code passed every test but you still did not understand why it works?