Using AI to Practise Interview Answers Well

There are two ways to use a language model in interview preparation. One of them makes you a noticeably better candidate in about three weeks. The other produces a folder of polished paragraphs and a person who cannot speak.

The dividing line is not which tool you use or how clever the prompt is. It is whether you speak before the model does. Everything below follows from that one rule.

The failure mode, described plainly

You paste in a question. You ask for a strong answer. You read it, agree with it, feel a small hit of progress, and move on to the next question. Repeat forty times and you have a document of competent, generic, slightly overwritten answers that all sound like the same person, because they were all written by the same thing.

Two problems with that. The obvious one is that reading is not speaking, and the gap between the two is where most candidates lose marks. The subtler one is that you have skipped the only part of preparation that actually transfers: the effort of working out, yourself, what you think about a hard situation and why.

Assessors are not comparing your answer to an ideal essay. They are watching a person reason in real time. A model can help you build that capacity. It cannot lend it to you.

The workflow

Five steps, roughly fifteen minutes per question. Do not reorder them, because the value is almost entirely in the sequence.

• Get the prompt and close everything else. Read it once, take the reading time you would get in a real station, then speak the answer out loud to a recorder. No notes, no restarts, no stopping when it goes badly.
• Transcribe it yourself, or paste in a rough transcript. Your actual words, filler included, not a tidied version of what you meant.
• Ask what you missed, not how to say it better. What considerations did this answer never raise? Whose perspective is absent? Where did I assert something instead of reasoning towards it?
• Argue back. If a suggestion does not fit the scenario or does not sound like you, say so and see whether it holds up. Half the time it will not, and noticing that is the skill.
• Answer again, spoken, immediately. Same question, same timer, incorporating at most two changes. Then stop and go to a different question.

Step five is the one people skip, and skipping it turns the whole exercise back into reading. The repeat is where a change moves from a note into your mouth.

Prompts that build you, prompts that replace you

Ask a model to write an answer and it will. Ask it to interrogate yours and it will do that too, usually better. The wording of the request decides which one you get.

Useful requests sound like: give me three considerations I did not mention. Ask me the follow up an assessor would ask about the weakest sentence in that answer. Steelman the position I argued against. Point out where I was vague without rewriting anything.

Damaging requests sound like: write a model answer. Improve this. Make it sound more confident. Each of those hands the work back to the machine, and what returns is a paragraph in a voice that is not yours, which you will then try to memorise and deliver in a room where you are being marked on being a person.

One good habit: never let a generated sentence into your preparation notes. If an idea is worth keeping, write it down in your own words, badly, from memory. The rewriting is the learning.

Using it for roleplay

A model will happily play the upset patient, the flatmate who has been drinking before shifts, the colleague who cut a corner. It is genuinely useful for this, particularly for the opening ninety seconds, which is where most people freeze. Our guide to MMI roleplay stations covers what those stations are testing, and the practice value here is repetition volume: you can run the same opening eight times in an evening, which no human volunteer will tolerate.

Two things to insist on. Tell it to stay in character and not to break into feedback mid scene, because a character that starts coaching you is no longer a scenario. And tell it to be realistically difficult rather than dramatic. Real actors do not escalate for entertainment, they hold a position and wait to see whether you notice.

Speak your side out loud even when you are typing to it. Typing a roleplay trains your fingers. The station is spoken.

The things to be careful about

Do not trust a model on the specifics of any Australian or New Zealand admissions process. Station counts, timings, formats and dates change between cycles, and a confident wrong answer about a particular school's structure is worse than no answer. Read our overview of what an MMI is in Australia and New Zealand for the general shape of things, and check the university's current admissions page for anything you are going to plan around.

Be equally sceptical about clinical or ethical detail. A model can produce a plausible sounding claim about guidelines or law. You are not being assessed on citation accuracy, so if a fact is doing structural work in your answer and you cannot verify it, take it out and reason without it. That is a better answer anyway.

And watch for homogenisation. If every candidate in the country runs their answers through the same tools, the answers converge, and the ones that stand out are the ones carrying something specific that only that person could say. Your job is to protect the specific bits: the actual ward, the actual argument you got wrong, the actual reason you kept turning up.

Making the practice feel real

Sitting in pyjamas typing at a chat window is not the same task as speaking to a stranger under a clock. Close the gap where it is cheap to do so. Stand up. Use a timer. Once or twice in the fortnight before, run a session in the clothes you have chosen, using our notes on what to wear to a medical interview, so that nothing about the day is novel.

If you want the timed and marked version of this rather than a self managed loop, MasterMed's live AI interviewer runs spoken MMI stations to a clock and scores them against a rubric, which removes the temptation to stop when an answer starts going wrong. The first speaking station is free on the trial, no credit card, and the trial never converts by itself.

The test for whether your AI practice is working is easy to apply. Two weeks in, can you answer a question you have never seen, out loud, without freezing? If yes, the tool is building you. If you can only produce good answers when you can see the screen, it has been writing for you, and you should change the order tonight.