Using AI as a Mock Interviewer, and Where It Misleads You
A chat model is a genuinely useful rehearsal partner for an interview, and it is useful for an unglamorous reason: it will ask you a hundred questions at six in the morning without getting bored, and it does not mind that your first three answers are bad. That is most of what rehearsal is. What it cannot do is assess you, and the gap between “it asked me things” and “it told me I was ready” is where people get hurt.
Used as a question generator and a talking-out-loud prompt, it earns its place. Used as a judge, it will tell you that an answer you invented was excellent.
Set it up as an interviewer, not an assistant
The default behaviour of these tools is to be helpful, which means writing your answers for you. You have to instruct that away.
Give it three inputs: the job posting, the resume you actually sent for that application, and a one-line statement of the role and level. Then the rules:
- Ask one question at a time and wait for my answer.
- Do not write, suggest, or improve my answers.
- After each answer, tell me only which parts were vague or unsupported.
- Stay in the role of an interviewer for this specific posting.
Then answer out loud, typing only a rough transcript. The point of speaking is that your mouth catches the sentences your eyes let through. An answer that reads fine and sounds like a press release is a discovery you can only make by hearing it.
If you find yourself typing polished paragraphs, you have stopped rehearsing and started drafting. That is the same failure mode as handing a session your resume and letting it steer instead of steering it yourself.
What it is actually good at
Producing the obvious questions in bulk. Most of an interview is predictable from the two documents in front of the interviewer, and a model reading those documents predicts it about as well as you would. It will find the short tenure, the unusual title, the tool listed once. That is the same list you get from reading your own page for the questions it invites, arrived at from the other direction, which makes it a decent cross-check on your own pass.
Follow-ups. Ask it to keep pushing on one answer for four turns. Real interviews go deeper than the first question, and depth is where unprepared answers fall apart. A model is tireless about “and how did you measure that?”
Naming vagueness. Instructed to point at unclear phrasing rather than fix it, it is reliable at flagging the places where you said something that could mean anything. That is a real service, because you cannot hear your own filler.
Rehearsing the same answer differently. Saying one story five ways, out loud, is how you stop reciting it. The model does not care that it is the fifth time.
Where it misleads you, specifically
It takes every claim at face value. It has no access to your record, so it cannot tell an accurate answer from a confident invention. Say you led a migration that you actually assisted on, and it will ask an interested follow-up about your leadership approach. A real interviewer might ask who else was in the room. The model cannot run the check that comparing a bullet against your own record runs, and it will never be the thing that catches an overstatement.
Its praise tracks fluency, not substance. Smooth, well-structured, jargon-appropriate answers get approving responses. So do smooth, well-structured, empty ones. If you optimise against its feedback you will end up more polished and no more prepared, and polish is not what the person across the table is short of information about.
It does not know the employer. It knows the posting text and whatever general patterns it has absorbed. It does not know that this team was reorganised last quarter, or that the person interviewing you built the system you are proposing to replace. This is the same limit as what a tool does not know about the employer you are writing to, and it means the questions it generates are the generic ones. Useful, but the specific ones will be the hard ones.
Its scoring is a plausible-looking rubric. Ask for a score out of ten and you will get one. It is a sentence-shaped object, not an assessment, and treating it as a readiness signal is how people walk into interviews believing they are prepared.
It produces script-shaped answers if you let it. Once you have read a model’s version of your answer, that phrasing is in your head, and under pressure you will reach for it. Recited answers are audible.
The one rule that keeps it honest
Never let it supply a fact. If it asks how much time the project saved and you do not know, the answer to type is “I do not know” — not a plausible figure, and not a request for a suggestion. Every gap you hit is an entry for your list of things to go and look up, which is exactly what a facts file is for.
Keep the transcript, and mine it for two things only: the questions you did not expect, and the facts you could not retrieve. Throw the model’s answers away.
What it cannot replace
One person who knows the field, asking you a question and then waiting. The waiting is the part no tool does — the silence that makes you add the detail you were going to leave out. If you can get one honest human rehearsal, do that, and use the model for the volume around it.
And when the format is a recording rather than a conversation, rehearsal changes shape again: there is no follow-up at all, which makes answering into a camera with no interviewer its own skill. A model can generate the prompts for that, but it cannot practise the part that is hard about it.