AI and Diagnostic Responsibility

The prompt usually arrives in the same shape. A decision support tool contributed to a clinical decision. The decision was wrong. A patient was harmed. Who is responsible?

Most candidates answer in four seconds: the doctor, because a doctor is always ultimately responsible. That is not wrong, and it is not an answer. It is the conclusion with the reasoning deleted, and the interviewer's next question is designed to find out whether you had any.

Separate the two things you are being asked

Responsibility questions in medicine contain two distinct things that students routinely fuse: who is answerable to the patient, and who or what caused the failure.

The first is about a duty to a person: someone has to explain what happened, apologise, and stay involved in putting it right. That is the treating clinician, and it does not distribute. You cannot tell a patient that the software owes them an explanation.

The second is about preventing the next one, and it distributes widely: the developer who trained the model, the regulator who approved it, the hospital that procured and deployed it, the person who decided how it appears on screen, the roster that put an exhausted clinician in front of it. Fixing the system requires looking at all of those. Facing the patient does not.

Say those two things apart and you have already given a better answer than the confident one liner, because you have explained why the doctor being answerable does not mean the doctor is the whole story.

The problem with the human in the loop

Everyone reaches for the same safeguard. The tool only advises, a clinician reviews it, so the clinician remains responsible. It sounds airtight and it has a well known weakness.

For oversight to be real, the reviewer needs three things: the time to actually review, enough information to disagree, and the standing to act on a disagreement. Strip any of the three and the human in the loop becomes a signature rather than a check.

• Time. If a tool exists because a department is under resourced, the reviewer is by definition rushed. The efficiency gain and the safety check are pulling against each other.
• Information. If the output is a number with no reasoning attached, the clinician can accept it or ignore it but cannot really evaluate it. Being unable to interrogate a recommendation is not the same as endorsing it.
• Standing. If overriding the tool means extra paperwork, or a conversation with someone senior, or a risk of being second guessed when you are wrong, most people will not override it. Culture decides how often the loop is exercised.

Holding someone accountable for oversight they were not resourced to perform is a familiar injustice in medicine, and it long predates algorithms. Making that connection is a strong move, because it shows you see the new problem as a version of an old one.

The two directions of error

Interviewers often present only one version: the clinician followed the tool and the tool was wrong. Bring up the other one yourself, because it is more interesting.

What if the clinician overrode a correct recommendation? If deviating from a tool starts to look like negligence, clinicians will stop deviating, and independent judgement quietly disappears. The mere existence of a recorded recommendation changes behaviour, because now every disagreement is documented and reviewable in hindsight. That is a subtle and real risk, and almost nobody mentions it.

A system that punishes only one kind of error will get one kind of behaviour, and it will not be the safe kind.

What you owe the patient

If the station has a roleplay component, the ethics stop being abstract. Someone has been harmed and is sitting in front of you asking what happened.

Do not hide behind the system. There is a version of this that sounds like blaming a computer, and it lands appallingly. The tool did not miss it. The service missed it, and you were part of the service. Say what happened in plain language, say what is being done to find out why, apologise properly, and do not speculate about fault before anyone has actually looked.

The tension worth naming out loud is that openness about how a decision was made is owed to the patient, and it can be genuinely hard to deliver when the answer involves a model nobody in the room can fully explain. Acknowledging that difficulty is honest. Pretending you could explain it would not be.

Consent and disclosure

A reasonable follow up asks whether patients should be told an algorithm is involved. Think about it in terms of what a patient would want to know in order to make a decision. Most people would want to know if a machine contributed to a diagnosis, particularly if they might want a second opinion, and being told after harm rather than before is what turns a defensible practice into a scandal.

The sensible line is proportionality. Nobody expects disclosure of every piece of software in a hospital, and a tool that materially shapes a diagnosis or a treatment recommendation is different from a scheduling system. Where a patient's decision could reasonably change, tell them.

Structuring it in a station

Split answerability from causation. Say who faces the patient. Then say why the systemic layer matters anyway. Then take one complication seriously, ideally the oversight problem or the override problem, and say what would need to be true for the safeguard to be genuine. Our guide to ethical stations in Australian MMIs sets out why that ordering, position then complication, reads as reasoning rather than as a list.

Panel interviewers tend to chase this topic further than MMI assessors have time for, so if you are facing a panel, prepare the second and third questions rather than the opening one. Our comparison of MMI and panel interviews in Australia explains the difference in how depth gets tested. Formats vary by university and change between cycles, so check the university's current admissions page for what you will face.

Things not to say

• The doctor is always responsible, full stop. True as a starting point, useless as an ending one.
• The company should be liable. Maybe, but you are not being asked to draft legislation, and confident claims about liability law will not survive one follow up.
• AI should never be used for diagnosis. Absolutist, and it ignores the patients currently waiting months for a report.
• Any invented figure about accuracy, approval status or a named product. Reason about mechanisms instead, and you cannot be caught out.

Getting it out of your head and into your mouth

This reasoning is not hard to follow on a page and it is easy to fumble under a timer, mostly because the two senses of responsibility are slippery when you are speaking quickly. Rehearse it against varied prompts from our collection of MMI interview questions used in Australia so the distinction becomes automatic rather than something you reconstruct each time.

MasterMed's live AI interviewer runs timed MMI stations and marks you against a rubric, so you can check whether the distinction actually survives being said out loud at speed. The first speaking station is free on the trial, no card, and the trial does not convert by itself.

The reason this topic keeps appearing is that it is unresolved in practice, not just in interviews. Nobody expects a school leaver to settle it. They expect you to notice that accountability is a structure rather than a single name, and to still know who has to walk into the room and say sorry.