Ask ChatGPT to run a mock interview and it will tell you that your answers are great.
That is the problem. Out of the box, every mainstream AI model is agreeable. It wants the conversation to go well. Ask it to interview you and it lobs soft questions, accepts your first answer, and rounds off with encouragement. You finish the session feeling prepared and you are not, because you have rehearsed against an interviewer that does not exist. No hiring manager nods along. They interrupt, they probe, they ask you to prove it.
Phase 4 of the Evidence-First approach fixes this the same way every other phase works: with structure. You do not ask AI to interview you. You instruct it, precisely, to interview you the way the difficult interviewer will.
Build the interviewer you’re afraid of
The mock interview prompt needs three things the casual version lacks: the decoded advert (so the questions come from what this employer actually cares about), your evidence bank and tailored CV (so the follow-ups probe your real claims), and an explicit instruction to push back.
You are interviewing me for the role described in the decoded requirements below. My CV and evidence bank entries are also below. Run a realistic interview, one question at a time, waiting for my answer before continuing.
Your behaviour: ask follow-up questions that probe for specifics whenever my answer is vague; challenge any claim that lacks evidence, numbers or a named outcome; ask at least one question about the weakest match between my evidence and the requirements; do not compliment my answers or reassure me.
After eight questions, stop and give me a blunt assessment: which answers would survive a sceptical interviewer, which would not, and the single answer most likely to lose me the role.
Two design choices matter here. “One question at a time, waiting for my answer” turns it into rehearsal rather than a reading exercise. And “do not compliment my answers” is doing more work than it appears to: it is the line that switches off the flattery and gets you the interviewer who is actually useful.
Notice also what this session really is. In the previous post, check four of the detection test asked whether you could defend every claim in your application for two minutes out loud. This is where you find out. Every claim in your CV and cover letter came from your evidence bank, so every claim is defensible in principle. The mock interview converts “defensible in principle” into “defended, out loud, under pressure, more than once”.
Rehearse the weak spot deliberately
One pass through is not preparation. The value is in the loop: run the interview, take the blunt assessment, strengthen the losing answer, run it again. If the model keeps flagging the same gap, that is not the model being difficult. That is the gap your real interviewer will find, surfaced early enough to do something about it, whether that is sharper framing of the evidence you have or an honest line about the experience you lack and how you would close it.
The one thing you do not do is ask the model to write better answers for you. The no-invention constraint applies to spoken words as much as written ones. Scripted answers collapse under the first unscripted follow-up. The model’s job is to pressure-test and sharpen what you can genuinely say, not to hand you lines.
When the interviewer is not human
The other half of Phase 4 is the mirror image: increasingly, the first interview is conducted by AI. Asynchronous video interviews scored by software, chatbot screeners, automated assessments. The state of play data covered how widespread automated screening has become, and if you are applying at any volume you will meet it.
Preparing for an AI interviewer is mostly the same discipline with the dial turned further towards structure. Software scores what it can parse: answer the question that was asked, front-load the direct answer, then support it with the specific evidence, and say the concrete things out loud, the system, the number, the outcome, rather than gesturing at them. A structured answer built on real evidence performs well with both audiences; the difference is that the human might rescue a rambling answer with a follow-up question, and the software will not. Rehearse for the AI interview exactly as above, but add a constraint to your mock prompt: answers of ninety seconds maximum, direct answer first.
And know that this territory is regulated. Where automated decision-making is involved, UK candidates have rights: to know it is happening, to request human review of a rejection, to challenge the decision. The next post in this series covers those rights in full, including the ICO’s expectations of employers, because most candidates have no idea how much of the process they are entitled to question.
The salary conversation
Phase 4 ends where the offer arrives, and AI has a role there too, in the same amplifier-not-author capacity. Use it to structure your research (what evidence would justify the top of the band for this role in this market) and then to rehearse the conversation the same way you rehearsed the interview: a negotiation partner instructed to push back, raise objections, and refuse the first ask. The negotiating position itself comes from your evidence bank; if the whole series has a single lesson, it is that the strongest thing you can bring to any stage of this process is a documented record of what you have actually done.
The line to hold
An interview answer is ready when it has survived an interviewer with no interest in being kind to you. Build that interviewer, run the loop, and walk into the real conversation having already met the hardest version of it.
Next in the series: your rights when AI screens you, what UK law requires of employers using automated decision-making, and how to use it.
Get the full framework
This post is part of a series drawn from my whitepaper, The Evidence-First Job Search: a complete framework for using AI across job search, applications and interviews, with a 15-prompt library, an evidence bank template, and a 30-day plan. It is free to download.
0 Comments