Say What You Fixed, Not What the Model Wrote

Give One Worked Example

Pick one task from the last few months where you used an AI assistant, and tell it in three beats: what you handed over, what came back, and the specific thing you changed and why. A reviewer can rate a decision. A claim to use AI "a lot" or "where it helps" rates nothing, because it names no decision at all.

The three beats, concretely:

• The task. One sentence: what you were trying to produce, and for whom.
• What came back. What the assistant actually gave you, including the part that was wrong, thin, or off for your situation.
• What you changed and why. The specific edit, and the thing outside the conversation you checked it against: a source, a colleague, a number that did not match.

The example can be small. A one-line formula that returned the wrong total, a paragraph that cited a policy that had since changed, a draft that assumed a context you did not have: each one gives an interviewer something to ask a real follow-up about, which is the entire point.

Have two examples ready, not one. Interviewers ask about whatever kind of work the role actually involves, and a story built around writing will not carry a question about a spreadsheet or a piece of code. Pick one case where the first output was mostly right and a small edit fixed it, and one where it was clearly wrong and you caught it before it went anywhere. Between the two you can answer either a light-edit question or a real-correction question without reaching for an example on the spot, which is where most answers to this question start to wander.

Why the Error You Caught Is the Strongest Move

The guidance interviewers are given for this question is built around one behavior, not around enthusiasm or fluency: whether you can say what a model got wrong and how you knew. What interview questions actually show how a candidate works with AI sets out the version employers are told to run: ask about one task, then work the candidate's own answer, what they typed first, what they refused, what they checked outside the chat.

In a 2023 Harvard Business School field experiment, consultants given GPT-4 on one task chosen to sit outside the model's capability were on average 19 percentage points less likely to reach the correct answer than a control group working without it. ¹ The failure mode was not laziness. It was a confident wrong answer on a task that looked like the ones the model handles well. Noticing that gap in the moment is the rare part, and it is what a worked example with a correction in it demonstrates.

Performing enthusiasm backfires for the same reason. A candidate who describes AI as central to everything they do, with no case where the output needed a fix, is telling an interviewer that either the work was never checked or the story has been smoothed over for the room. Neither reading helps you. The version that holds up under questioning is smaller and more specific: one task, one output, one thing you changed, told plainly.

What Not to Claim

Do not describe a tool or a workflow you cannot walk through this way. An interviewer who has run this question before will ask what you checked, and a claim built on nothing specific runs out of road in the second follow-up.

Do not deny using AI on work where you did, either. In Greenhouse's 2026 survey of 373 hiring managers in the UK, Ireland and Germany, 37% named detecting AI-generated or heavily AI-assisted applications among their top hiring challenges, and 31% said they were running more in-person, on-site interviews, though the report does not connect those two answers. ² A live stage puts a person in front of your work who can ask about it, which is where a false denial has the furthest to fall. The nearest evidence is not from hiring at all: in a vignette experiment about a freelance tax advisor, participants trusted him least when a third party exposed his undisclosed AI use, and more when he had disclosed it himself. ³ The ordering is what carries over, not the size of it.

That asymmetry is the reason a flat "I don't really use it" is the wrong move if it isn't true, and the right one if it is. If your honest answer is that you barely use AI at all, that is a complete answer of a different shape, and it deserves its own account rather than an invented one. If you barely use AI, say so and bring the judgment covers that case directly, including what to say when the posting asked for fluency you do not have yet.

How to Handle a Follow-Up You Weren't Ready For

Stay inside the same example rather than reaching for a new one, and answer the narrow question you were actually asked. Expect one or two follow-ups: what specifically told you the output was wrong, what you would check if the same thing happened tomorrow, or what you would have done if you had not caught it.

If you cannot remember an exact figure or the precise wording of what the model produced, say so and describe what you would check now rather than filling the gap with something invented. A guess dressed as a memory is easy to spot and hard to recover from, while "I'd want to see the underlying number before I trusted that line" is a real answer that moves the conversation forward.

Using AI at work is not a fringe habit you are confessing to: in nationally representative US surveys run in late 2024, 23% of employed people had used generative AI for work at least once in the previous week. ⁴ The question was never whether you used it. It is whether you can say, in your own words, what you did with what it gave you, on a task specific enough that a follow-up question has somewhere to go.

See how it works >