There is a quiet fear running through clinical teaching right now, and it is rarely said out loud because it sounds faintly ridiculous. It goes: the student in front of me has a tool on their phone that will produce a better differential than I will, faster, and without the coffee. On what grounds, exactly, am I supervising them?
It sounds ridiculous because it is built on a false premise, and the premise is worth taking apart slowly, because the rest of the argument about AI in education leans on it. The premise is that supervision was ever about knowing the answer first.
What a supervisor was actually for
Think about the best supervisor you ever had. I would bet that what made them good was not that they always knew the diagnosis before you did. Sometimes they did. More often, what they did was ask the question that exposed where your reasoning had gone soft: Why that and not the other thing? What would make you change your mind? What are you going to check before you act on this? The value was never in the answer they were holding back. It was in the pressure they applied to the path you took to yours.
That pressure does not become obsolete when a machine joins the conversation. It becomes the whole job.
At AMEE in Vienna this summer, a session on clinical reasoning in the era of large language models made this concrete in a way I found useful. It drew on a supervision framework published last year in the New England Journal of Medicine by Abdulnour, Gin and Boscardin, which adapts an older teaching structure into what they call DEFT-AI. Stripped to its questions, it asks the learner five things:
- Diagnosis and discussion. What is your assessment — and what did you ask the AI?
- Evidence. What supports your conclusion? Why should this tool be trusted on this particular question?
- Feedback. Where did your reasoning and the tool’s output part ways, and why?
- Teaching. What do you still need to verify before you act, and how?
- AI engagement. Next time you face this kind of task, how should you use the tool?
Look at what is missing from that list. None of the questions requires the supervisor to out-diagnose the model. Every one of them is aimed at the learner’s judgement process — at the prompt they wrote, the trust they extended, the divergence they noticed or failed to notice. A supervisor who could not beat the tool on a quiz can still ask all five, and can still tell a good answer to them from a bad one.
That, for anyone responsible for faculty development, is the most important sentence in this whole debate. The psychological barrier to teachers engaging with AI is not technical. It is the sense of having been demoted. The demotion is imaginary.
Permission by evidence, not by calendar
The second idea from that session is less comforting and more interesting. The usual institutional question is may students use AI?, answered with a policy that varies by year: nothing in the first year, some in the clinical years, whatever they like after graduation. That is the wrong grain.
The better question is: which part of the task may the AI enter, given what this learner has already shown they can do? The session sketched a ladder. Before a learner has demonstrated much, the AI generates practice — case variations, extra questions. Once they can frame a clinical question on their own, they may consult it. Once they can write a note independently, ambient documentation becomes reasonable. Later still, the tool can support evidence retrieval and drafting — but the work of fitting all that to this patient, and deciding what to do, stays with the human.
I would not treat that particular ladder as settled; it is a proposal, not a validated curriculum. But the principle behind it is familiar to anyone who knows how entrustment works in postgraduate training. We do not let a resident do something because it is March of their second year. We let them do it because someone has watched them do the smaller version well.
Where the two ideas meet
Here is the part that I think the session implied without quite saying, and it is the reason I wanted to write this at all.
Graduated autonomy has an obvious weakness. It depends on continuously knowing what a learner has demonstrated, and that knowledge is expensive. Without a way to generate it, every well-meant ladder collapses back into the calendar, because the calendar is free.
The supervision conversation is the way to generate it. Every time a supervisor asks where did you and the tool diverge?, the answer is a data point about whether this learner can hold their own judgement against a fluent machine. Every time they ask what do you still need to verify?, they learn whether the learner knows the limits of what they were handed. The five questions are not a separate activity bolted on to entrustment. They are the instrument entrustment was missing.
And they point the supervisor at exactly the place where they add most. Models are unusually strong at diagnostic reasoning, which tends to come with a clear right answer to be trained against. They are weaker at management — the situated, negotiable business of what to do for this person, in this setting, with these constraints. The characteristic failure there is not a dangerous recommendation stated boldly. It is an omission: the thing that should have been said and was not. A supervisor is badly placed to compete with a model on the first. On the second, a supervisor who asks what did it not mention? is doing something the model cannot do for itself.
What I am not claiming
Two limits, stated plainly.
First, the neighbouring principle — make learners commit to their own reasoning before they consult the tool — sounds as if it should rescue all of this, and in at least some experiments it did not: people who reasoned first still did worse with the model than the model did alone. I wrote about that result earlier this month. The supervision questions are a way of looking at the judgement process; they are not proof that looking improves it.
Second, conversations cost time, and clinical teaching time is already the scarcest thing in the building. Asking five questions per AI-assisted case is not free, and I have not seen anyone show what it costs at scale.
But the fear I started with does not depend on either of those. It depends on the idea that teachers were only ever as valuable as the answers they held. They were not, and the arrival of a machine that holds answers more cheaply is, if anything, a chance to say so.
Reference: Abdulnour, R.-E. E., Gin, B., & Boscardin, C. K. (2025). Educational Strategies for Clinical Supervision of Artificial Intelligence Use. New England Journal of Medicine, 393(8), 786–797.