Where AI Interview Practice Falls Short

We build an AI interviewer, which means we spend more time than most people looking at what it gets wrong. It is a genuinely good instrument for volume, consistency and honest scoring. It is not a substitute for every part of preparation, and pretending otherwise would cost you an offer, not us.

So here are the gaps, specifically, and what to do about each one. None of them require a coach or a large amount of money. Most require one other human being and about two hours in total.

Gap one: a person who is actually affected by you

A model can play distressed. It cannot be distressed, and some part of you knows that. In a roleplay station the pressure that changes your behaviour is the sense that a real person in front of you is having a bad time and it is partly your fault. That pressure is what makes candidates rush, over explain, or reach for a solution before the person has finished being upset.

You cannot rehearse your response to that feeling if you never feel it. Practising with a model gets your structure clean and leaves the emotional load untested.

Cover it by running roleplay with a human at least three or four times before your first interview. It does not need to be another applicant. A sibling with a scenario card and instructions to stay upset and not help you is enough, and in some ways is better, because they will not soften out of collegial sympathy.

Gap two: being looked at

Nothing about talking to software reproduces the specific unpleasantness of two assessors writing while you speak, or an examiner whose face gives you absolutely nothing for four minutes. That is a physical event. Your hands change, your breathing changes, your pace changes, and the answer that was fluent alone in your room becomes eighty per cent of itself.

Cover it deliberately rather than hoping. Ask two people to sit in front of you, say nothing, take notes on paper, and not smile. Run three stations like that. It is a horrible twenty minutes and it is worth more than another ten solo repetitions, because it is the only way to find out what your body does when it is watched.

Gap three: it does not know your school or this cycle

Any model, ours included, is working from general patterns rather than from a live feed of Australian and New Zealand admissions arrangements. Station numbers, timings, interview windows and formats change between cycles and between schools. Use our overview of what an MMI is in Australia and New Zealand for the general shape, and then check the university's current admissions page for anything you are actually planning around. If a tool tells you a specific school runs a specific number of stations, treat that as a guess until the university says so.

The same applies to anyone promising real or past questions. Nobody has them, and a bank of allegedly leaked prompts is either invented or a breach of someone's agreement. Practising against invented prompts is fine and normal. Believing they are the actual paper is how people end up thrown by the first station.

Gap four: it is patient in ways a panel is not

Software waits. It does not sigh, glance at the clock, or interrupt to move you along. Real interviewers do all three, sometimes without meaning anything by it, and a candidate who has only practised in a patient environment reads those signals as failure and loses the next thirty seconds to panic.

Cover it by asking a practice partner to be mildly rude on purpose, once. Have them cut you off at ninety seconds, ask a follow up that ignores what you just said, and look unimpressed. Then answer anyway. The point is not to enjoy it. The point is that the first time it happens should not be in a scored station.

Gap five: it cannot tell you what you are like

A rubric score tells you whether an answer covered the ground. It does not tell you that you come across as slightly dismissive when you disagree, or that your face does something odd when you are buying time, or that you are warmest in the first thirty seconds and then go formal. Those impressions are exactly what a panel forms, and only a human who watched you can report them.

Ask one person you trust a direct question after they have watched three of your stations: what was I like to sit opposite? Not how was the answer. That question, asked once, is often the single most useful piece of feedback in a whole season.

Gap six: the mechanics of the actual day

Practice sessions do not replicate the logistics, and the logistics are where avoidable damage happens. Online interviews bring their own set of problems, from lighting and eye line to what happens when the connection drops, which we cover in our comparison of virtual and in person MMIs. In person days bring travel, waiting rooms and two hours of standing around with other nervous people.

Handle these once, on paper, weeks out. Decide the clothes and wear them for a full session, which our guide to what to wear to a medical interview will let you settle in ten minutes. Test the camera setup in the actual room at the actual time of day. Print the address. None of this makes you a better candidate, and all of it stops you arriving already depleted.

What it is genuinely better at

Being fair about the limits also means being clear about where the machine wins, because it wins in the places that consume most of your preparation time.

• Volume. Twelve stations on a Tuesday night with no scheduling, no favours owed and no need for anyone else to be free.
• Consistency. It applies the same criteria to your twentieth answer as your first, which no tired friend does at ten at night.
• Honesty without social cost. It has no reason to soften a score, and you have no reason to be embarrassed in front of it.
• Coverage. It will keep handing you the station types you have been avoiding, which is precisely what a friendly practice partner will not do.
• Repetition without shame. You can fail the same opening six times in a row, which is the fastest way to fix an opening.

The split that makes sense

Use the machine for reps and scoring, and reserve human sessions for the things only humans provide. In practice that looks like most of your weekly volume on timed stations, plus roughly one human session a fortnight built around roleplay, silent observers and a straight answer to what am I like to sit opposite. MasterMed's live AI interviewer covers the first half of that, with timed stations and rubric marking, and the first speaking station is free on the trial, no card, and the trial never converts by itself.

The candidates who do best are rarely the ones who found the perfect tool. They are the ones who worked out what their preparation was not testing and then went and tested it, usually by asking one person for one uncomfortable hour.