A Generated Answer Is Fluent. The Detail Is What Gets Scored.

What Does a Generator Actually Produce?

A well-formed answer to a generic version of your question, in the shape a rubric expects: situation, task, action, result, in language that reads as competent. That part genuinely works, and it works on the first try. What arrives with it is a plausible stand-in for the details. The model has your prompt and the job posting and no access to the Tuesday the release broke and you were the one who noticed.

The rating scale is built around exactly the material the model does not hold. In a meta-analysis comparing the two structured formats, past-behaviour questions scored with descriptively anchored scales produced a substantially higher mean validity estimate than situational questions with the same scales, .63 against .47 ². Those are corrected 2002 estimates resting on range-restriction assumptions later shown to be generous, so the comparison between formats is the part that survives, not the numbers themselves.

The format also holds up better against bluffing.

An experiment in a 2007 paper on interview faking found past-behaviour questions more resistant to faking than situational ones, and also found that follow-up questioning increased faking rather than damping it ¹. More resistant is relative, not immune. The study ran with undergraduates in person, and a one-way video interview usually has no follow-up at all.

A generator can invent the event, and invention is the part that tracked with lower ratings.

The study that gets closest ran as a mock asynchronous video interview taken by 582 people across ten countries, where honest tactics such as self-promotion were positively associated with rated performance ³. Deceptive ones, extensive image creation among them, were negatively associated with it. Human raters, no job at stake, self-reported tactic use. It shows an association, not a rule inside any scoring model.

Why Does the Polished Answer Lose Its Value?

Because it stops carrying information once everyone has it. When an AI cover-letter tool arrived on a large online labour platform, letters matched the job posts more closely and callback rates rose for a while. Then the correlation between how well a letter matched the posting and whether it got a callback fell by 51 percent ⁴. Employers shifted weight onto other signals, including prior work history, which a generator has no way to supply.

That study is an arXiv preprint using difference-in-differences on one unnamed marketplace, so it is correlational and it is not corporate hiring, and the authors report the callback lift itself as fragile and short-lived. The 51 percent is a fall in a correlation between text-matching and callbacks, not a fall in anyone's callback rate.

Nobody has measured the same effect for a spoken answer, but the logic carries.

A handful of models answer the same standard questions for many people preparing at once, so the phrasing that feels sharp to you may be phrasing the panel has already heard. Why every candidate gives the same polished STAR answer is written for the interviewer sitting across from it. Hiring teams are being pointed at the follow-up questions that expose whether somebody understands the answer they just gave.

Rehearsing is not the problem; sounding like everybody else is.

Why your STAR answers sound like everyone else's covers the same ground from your side of the table, including why an answer can be word-perfect, correctly structured and still land as nothing in particular.

Where Does the Live Version Fall Apart?

In the interview itself, where generated text has to be read aloud at conversational speed while somebody watches. The two behaviours have already separated in the survey data. In a UK early-careers survey, 47 percent of respondents said they used AI to practise or prepare for interviews in 2026, against 29 percent a year earlier. Use during the interviews themselves sat at roughly 5 percent ⁵.

Those are self-reported figures from a self-selected sample of 4,990 UK respondents, and all three percentages were read off a chart in the report, so hold them loosely ⁵. The direction, and the fact that use during interviews stayed far lower, are the durable parts.

Employers are watching the live version specifically. In the Institute of Student Employers' 2025 survey, using AI during interviews without disclosure or permission was the most commonly reported form of candidate misconduct, named by 61 percent of member employers ⁶. That figure counts what employers say they suspected or identified, with no verification step behind it, and ISE members are UK graduate and apprentice recruiters weighted towards large organisations. It is a count of suspicion rather than a measurement of what candidates did.

What a fed answer does to your own answer is the more useful question.

Read text tends to arrive at the wrong tempo, put its emphasis in the wrong place, and answer the question that was typed rather than the one that was asked. It also carries nothing you did. The follow-up asks what you would have done differently, and the follow-up is where the material has to be yours or there is nothing true to say.

No draft holds what you did, however many follow-ups it answers.

At least one live-answer vendor puts the rule in its own paperwork. ParakeetAI's terms of service prohibit using the service to gain an unauthorized advantage in any interview where AI assistance is prohibited or has to be disclosed ⁷. The same site was marketing undetectability on the day those terms were read. The question of what is permitted here sits with you either way.

Use the Draft, Then Rewrite Every Line

Generate the draft if it helps you start, keep the structure, and throw out every sentence that could belong to somebody else. The one measured signal here points the same way. On the freelancing platform in that cover-letter study, over 75 percent of AI-drafted letters were submitted within a minute of clicking generate. Longer editing time was associated with a higher chance of winning the work ⁸.

The authors say plainly that this is not causal. The decision to spend longer editing is unlikely to be random, and people who edit more may simply be more motivated or more skilled. The base rates are small, so even the associated gain is a small change in absolute terms.

On that platform, almost nobody spent the extra minutes.

• Replace the invented example with a real one: the date, the constraint, the part that went wrong.
• Cut any line that would be true for any applicant to any job.
• Give the number you actually remember, and say you are estimating when you are.
• Rehearse the follow-up rather than the answer: what you would do differently now, and what it cost.

Then read what the employer has published, because some of them are specific. Anthropic sets a rule per stage rather than a blanket ban. It asks candidates to draft the application themselves and refine it with Claude, encourages Claude for interview preparation, and rules out AI on take-homes and live interviews unless it says otherwise ⁹. That page is dated 10 July 2025 and states no detection method and no penalty.

Preparing with ChatGPT is allowed and using it live is not covers where that line usually sits, and what to ask when an invitation says nothing.

See how it works >