Preparing With ChatGPT Is Allowed. Using It Live Is Not.

Where Does Preparation End and Live Help Begin?

The line sits at the moment the interviewer starts listening. Before that moment, you can rehearse with anyone or anything, because nobody in the room is being asked to judge a conversation that hasn't happened yet. After it, an assistant drafting or feeding you lines turns the interview into a transcript of the model's reasoning wearing your name. That is a claim about authorship, not a productivity choice.

You are left guessing at it because the rule is often unpublished. In the Institute of Student Employers' 2025 survey of UK graduate and apprentice employers, using AI during interviews without disclosure or permission was the most commonly reported form of candidate misconduct, named by 61% of them, while 45% had given candidates no guidance at all on when AI use is or isn't appropriate ¹. That gap, permission assumed on one side and never stated on the other, is where most of the anxiety about this question comes from.

Where an employer does publish a rule, follow it over anything here. Should candidates use AI during the interview, or should it be banned? works through the same line from the other chair: the argument put to employers there is to permit in each round exactly what the job itself permits, to close only the step that genuinely has to be done unaided, and to put that rule in the invitation rather than issue a blanket ban. A ban and an allowance are both defensible on that reasoning, which is why the rule you are given can differ by role.

Notice what the line is not. It is not about how much you relied on a model, how many hours you spent with it open, or whether a career coach would approve of the method. Two candidates who prepared identically, one with a chatbot and one with a human coach charging by the hour, are in the same position once the call starts. The thing being judged live is whether the words coming out of you are yours, not how you got ready to say them.

How to Prep Without Sounding Generated

Prep works when you use a model to interrogate your own material instead of asking it to produce material of its own. Paste in the question, then your real answer, and ask what a skeptical interviewer would push on next. Ask it to name the weakest part of your story before you walk in, not to write a stronger one.

Never memorize a sentence it drafted for you. A phrase that came from the model rather than from your own memory of the event is the one an interviewer's second question finds, because you can defend a decision you actually made and you cannot defend the wording of a paragraph you didn't write. Three uses hold up:

• Ask it to argue against you. Feed it your answer and ask what a hiring manager would doubt about it. The doubt is more useful than a rewrite.
• Ask for the follow-up, not the answer. Have it generate three probing questions on your story rather than a version of the story itself.
• Use it to structure, never to author. Turning a messy memory into a Situation, Task, Action, Result outline is organizing what already happened. Writing the sentences that describe it is a different job, and it should stay yours.

A model can produce a fluent, well-organized answer in seconds, and that speed is exactly the trap. In a 2023 experiment, 444 professionals given ChatGPT finished occupation-specific writing tasks 37% faster than a control group, and graders scored their output higher ². Neither of those is what an interview measures. Fast and polished is a worse trade in the room than slow and yours.

What Counts as Live Help?

Live help is anything that puts the model's words in your mouth in real time: a copilot overlay reading the interviewer's question and feeding you a response, a browser tab you're quietly reading from, a second device transcribing and drafting while you talk. It does not include a notes document you wrote yourself the night before, even if a model helped you draft it.

The separator is timing: whether the sentence existed before you sat down, or is being generated while you're supposedly speaking for yourself. The reason to hold that line has little to do with getting caught. If a live assistant is found out later in the process, on a reference check or a second-round conversation with someone who compares notes, the cost is worse than if the same use had been disclosed up front. In one online experiment on AI disclosure, set between a professional and his client rather than in hiring, trust was rated lowest when a third party exposed undisclosed AI use, below the rating for the same use disclosed by the person himself ³. That ordering is what carries across; the size of it in an interview is not something anyone has measured. Silence isn't neutral. It's a bet that nobody ever compares notes.

Check the Employer's Policy First

Check the careers page and the invitation email before assuming either extreme. At least one employer publishes an actual, stage-by-stage rule. Anthropic's candidate guidance, dated July 2025, asks applicants to draft the written application themselves and use its assistant only to refine it, encourages using it to prepare for interviews, and reserves a flat no-assistant rule for the live round and any take-home unless the recruiter says otherwise ⁴.

That is one company's rule on one date, and the company says it expects to revise it. It is not a norm, but it shows what a careers page can actually tell a candidate when it bothers to. Where nothing is published, ask the recruiter directly what's expected for that round. It's an ordinary logistics question, not a confession, and it puts the rule on the record instead of on a guess.

If you get an answer, write it down somewhere you'll actually check before the call, because "I asked once" is easy to forget under nerves. If you don't get a clear answer, default to the version in this article: prep openly, arrive unaided. The same ask-first habit is worth carrying into the rest of your prep. How much you can lean on a model to research a company before the interview runs into the identical question of where a fast answer stops being a safe one.

See how it works >