AI in Medicine: Framing an Answer That Is Not Naive

There are two stock answers to AI questions and interviewers have heard both several hundred times. One is breathless: it will transform diagnosis, catch what humans miss, free doctors to spend more time with patients. The other is defensive: a machine cannot show empathy, medicine is fundamentally human, technology will never replace a doctor.

The problem with both is not that they are wrong. It is that neither one required any thinking, and the assessor is marking your thinking.

First move: refuse the word

AI in medicine is not one thing, and treating it as one thing is what forces you into a slogan. Break it apart in your first fifteen seconds and the rest of the answer gets easier.

• Pattern recognition on images and signals. Reading scans, screening slides, flagging rhythms. Narrow, testable, and the area with the most established use.
• Administrative and documentation tools. Note taking, letters, coding, rostering. Unglamorous, and probably where clinicians feel the change first.
• Risk prediction and triage. Estimating who will deteriorate, who will not attend, who should be seen first. The place where bias does the most damage, because it allocates attention.
• Patient facing tools. Symptom checkers and chatbots people use before they ever reach a clinician, often instead of reaching one.

Those four raise different questions. Bias in a rostering tool is an annoyance. Bias in a triage model changes who gets care. Saying that sentence alone puts you ahead of the field, because it demonstrates that you know the risks are not uniform.

Second move: ask what it is being compared to

This is the most useful habit in the whole topic. Hype compares AI to perfection and finds it wanting. Fear compares AI to an idealised doctor with unlimited time. Neither comparison exists.

The real comparison is to current practice: a tired registrar at three in the morning, a scan sitting in a queue for a week, a rural patient who has no specialist to see at all. A tool does not have to be flawless to be worth having. It has to be better than the alternative that is actually available, and safe enough that its failures are visible.

That last clause matters. Human error and machine error are distributed differently. A tired clinician makes scattered mistakes. A flawed model makes the same mistake consistently, at scale, on the same kind of patient, and nobody notices because it is confident every time.

Third move: name the concrete failure modes

Generic caution about ethics scores nothing. Specific failure modes score well. Three are worth knowing properly.

Training data bias

A model learns from the patients it was shown. If a population was under represented in that data, performance on those patients can be worse while overall accuracy still looks excellent. Aggregate numbers hide subgroup failure, which is precisely why an impressive headline figure is not reassurance.

Automation bias

This is the one most candidates have never heard of, and it is the strongest thing you can bring. Humans overtrust automated output, especially when busy. A tool designed as a second opinion quietly becomes the first, and the clinician's independent judgement erodes without anyone deciding it should. The safeguard of a human in the loop only works if the human is genuinely still thinking.

Deskilling

If trainees learn alongside a tool that does the pattern recognition, do they develop the skill to check it? This is a slow risk with no obvious moment of failure, and it is a genuinely open question rather than a settled objection. Saying so is fine. Pretending to know the answer is not.

Do not oversell the empathy line

Almost every candidate reaches for the same reassurance: machines cannot be empathetic. Be careful, because interviewers increasingly push on it. Patients sometimes disclose more to a screen than to a person, precisely because a screen does not judge. Written information from a machine can be clearer and more patient than an exhausted clinician at the end of a shift.

The more defensible version is about responsibility rather than feeling. A clinician can be accountable for a decision, can be asked why, can be sued, can sit with someone through the consequences and can change their mind because of who is in front of them. That is not warmth. It is a structure of answerability that a tool does not have, and it does not collapse the moment someone points out that a chatbot sounds kind.

Putting it together under time

In a short station you cannot do all of that, so pick a spine: split the category, name the realistic comparison, choose one failure mode and develop it properly, then land somewhere. Our overview of what an MMI is in Australia and New Zealand explains why depth on one point beats breadth across five when the clock is running.

In a panel you will get room to expand, and you should expect the interviewer to test whether your position is genuinely yours. The differences between the formats are covered in our comparison of MMI and panel interviews in Australia. Interview formats vary by university and change between cycles, so check the university's current admissions page rather than assuming.

Follow ups that catch people

• Would you tell a patient that a tool contributed to their diagnosis? Yes, and think about why: it is their information, and it affects whether they can meaningfully question the result.
• If a system outperforms clinicians on average, is it unethical not to use it? A sharp question. Average performance is not the whole picture, but you should feel the force of it rather than dodging.
• What if you disagree with the tool? Your judgement stands, you document your reasoning, and you should be able to explain the disagreement to a patient without hiding behind either the machine or yourself.

The accountability strand goes deep quickly, and it is worth preparing as its own topic. The general approach to holding a position under escalating pressure is covered in our guide to ethical stations in Australian MMIs.

A note on honesty

Do not quote accuracy figures or name studies you have not read. This field moves fast, claims get repeated inaccurately, and an interviewer who works in the area will spot a half remembered number instantly. Reasoning about mechanisms cannot be caught out in the same way, and it is more persuasive anyway.

This is a topic where written preparation misleads you, because the structure looks clean on paper and comes out as a lecture in the room. MasterMed's live AI interviewer runs timed MMI stations and marks you against a rubric, which is the quickest way to hear whether you sound thoughtful or rehearsed. The first speaking station is free on the trial, no card, and the trial does not convert by itself.

The test is not whether you are optimistic or worried. It is whether you can hold a specific view about a specific application, know what would change it, and resist the pull of the two comfortable scripts everyone else is using.