What an AI copilot costs you in a real interview

There is a category of product now that listens to your interview and puts a suggested answer on your screen within a few seconds. Final Round AI, Verve AI and several others do exactly this, openly, as the main feature.

The obvious question is whether it works. I think that is the wrong question, and I want to answer a more useful one: what happens in the room when somebody uses one.

I ran final-round interviews for ten years. I have not caught anyone red-handed, and that is rather the point — you do not catch people, you notice something is off and the interview quietly goes worse.

What an interviewer notices

Not the tool. The rhythm.

The pause before every answer is the same length. People do not think in fixed intervals. A question you know cold gets an answer in half a second; a question you have to reconstruct gets four seconds and a false start. A steady two-second gap before every response, including the easy ones, is strange in a way that registers before it is understood.

The answer is better than the follow-up. This is the big one. A generated answer is complete, well-structured and slightly generic — a good textbook paragraph. Then the interviewer asks "why that and not the other option in your case?" and the second answer is thinner than the first. In an unassisted candidate the follow-up is usually stronger, because that is where they stop reciting and start talking about what they actually did.

Eye movement. Everyone glances away while thinking. Reading is a different motion, at a different rhythm, and on a video call it is more visible than people expect.

None of this produces an accusation. It produces doubt, and doubt in a hiring decision resolves as a no. Nobody writes "suspected AI use" in the notes; they write "seemed to know the theory but I could not get to what he had actually done", and that is fatal in a way you never get feedback on.

The part that is worse than being caught

Suppose it works perfectly. You get the offer.

Now you are three weeks into a job where the expectations were set by someone who answers system design questions cleanly and reasons out loud about trade-offs. That person was not you, and the gap shows up in the first design review, in front of the team, with no copilot and no undo.

This is not a moral argument. It is a practical one: the interview is a filter in both directions, and beating it in the direction you cannot sustain is an expensive way to win.

I have watched a mis-levelled hire play out from the manager's side. The person was not incompetent. They had been hired for a level they could not hold, everyone spent six months finding that out, and the ending was bad for them more than for anyone else.

Where these tools are genuinely useful

I want to be fair to the category, because the underlying technology is not the problem — the timing is.

Before the interview, they are excellent. Practising out loud against generated questions, hearing your own answer played back, finding out that the explanation which is clear in your head takes ninety seconds and three restarts out loud — that is real preparation, and it is exactly what most candidates skip.

For non-native speakers, the accommodation argument is real. Interviewing in a second language is a genuine handicap unrelated to competence, and having help finding a word is not the same as being fed an answer. The line I would draw: support for how you say it is fine, supply of what you say is not.

In a live interview, the trade is bad. You are exchanging a small chance of a job you cannot hold for a real chance of an interview that goes flat for reasons nobody will explain to you.

What to do instead, concretely

The reason people reach for a copilot is fear of blanking on a question they half-know. That fear is reasonable and there is a better answer to it.

Practise out loud, not in your head. In your head every answer is fluent, because your head fills gaps automatically. This is the single highest-return thing most candidates never do.

Prepare four stories, not forty facts. Most behavioural questions are the same four or five questions wearing different clothes. Four pieces of work you can describe in ninety seconds each — problem, what you did, what changed — cover most of a round.

Learn to say "I do not know" well. "I have not worked with that directly. My guess would be X, because Y — is that close?" reads as senior. Every interviewer has heard confident nonsense and remembers who did it.

If you want to rehearse against questions for your role and level, we keep a set per specialisation on the questions pages, and practising them out loud costs nothing.

The thing worth remembering

The interviewer is not trying to establish whether you know the material. They are trying to establish what you would be like to work with when something breaks.

A copilot can answer the first question. It cannot answer the second, and the second is the one being asked.