The Canadian data contain a request, stated with unusual clarity: in Ontario survey findings, 91.6% of responding medical students recognised inaccuracy risks in generative AI, 78.9% recognised bias risks, and 75.9% supported formal incorporation into the curriculum as a resource or teaching topic. Read together, the three numbers dissolve the usual framing of students as naive enthusiasts needing restraint: this cohort is sceptical, fluent and asking to be taught properly, and the gap it names is specific, not "should AI be in the curriculum" but what belongs in it beyond the prompting workshops that currently pass for coverage.
What the requested curriculum should contain
Six components, each answering a risk the students already recognise. Model limitations as working knowledge: plausibility engines, stylistic confidence, the characteristic failure modes known by feel rather than by slogan, because recognising a hazard in the abstract and catching it in a fluent paragraph are different skills. Source verification as procedure: the Canadian source map, national and provincial guidance, drug monographs, and the timed workflow that makes checking cheap enough to actually happen. Indigenous and population-health contexts: a distinctively Canadian obligation, since models trained on globally skewed data will under-represent exactly the populations Canadian curricula have spent a decade centring, and AI literacy that omits this has taught appraisal while modelling the bias it warned about. Confidentiality in practice: the never-paste rules with placement-realistic scenarios, letters, images, distinctive cases, and the reasoning that name-removal is not anonymisation. Documentation and disclosure: what gets declared, where, in what form, so honesty is specified rather than improvised. And human accountability: the supervision skill, checking AI work as an assessable capability, which is the component clinical practice will examine whether or not medical school does.
Beyond prompting, deliberately
The list's negative space matters as much: prompt technique appears nowhere as a pillar, because it is the perishable layer, interface-specific, obsolete on each product cycle, and already self-taught by a cohort using these tools weekly. A curriculum hour spent on prompt patterns is an hour taken from appraisal and supervision, the durable capabilities, and the allocation choice is the difference between teaching this year's software and this generation's judgement. The same logic governs assessment: examining prompt fluency assesses tool operation; examining the critique of a deliberately flawed AI output, or the verification of a claim against Canadian guidance under time, assesses the thing the curriculum exists for, and both are buildable into existing OSCE and written formats without new machinery.
Where platforms end and curricula begin
An honest boundary statement, necessary because vendors, ourselves included, benefit from blurring it. Learning platforms can host the practice: grounded question-answering where verification is one click, tutoring that enforces attempt-first discipline, exam-mapped retrieval with spaced return, and iatroX positions itself as exactly that environment for MCCQE-track and internationally mobile Canadian students. What no platform can supply is the curriculum's own work: the appraisal standards, the population-health framing, the local governance, the assessment of supervision skill, and the professional formation that makes disclosure normal. A school that outsources those to any product has not closed the gap the students named; it has relabelled it. The 75.9% were asking their faculties, not their app stores, and the request is fillable with existing educational technology of the oldest kind: decided content, taught well, assessed honestly.
Frequently asked questions
How much curricular time does this actually need?
Less than feared: the six components fit a short longitudinal thread, a few hours per year with placement-integrated practice, provided assessment touches it; unassessed AI content of any volume rounds to zero.
Should the curriculum recommend specific tools?
Categories and criteria over brands: grounded versus generic, verification cost, jurisdiction fit, with worked examples that will date; endorsing named products converts education into procurement and ages badly.
What can students do while their school builds this?
Run the six components as a self-curriculum, most of it is procedure and habit, and feed the demand signal back formally; the survey suggests three-quarters of the room will co-sign the request.
Which component should schools build first?
Verification-as-procedure, because it is cheap, assessable and upstream of everything else: a cohort that checks claims against Canadian sources by reflex has converted the risk awareness the survey measured into behaviour, and every other component teaches faster on top of it.
How should curricula handle the pace of tool change?
By teaching at the durable layer and dating the perishable one: mechanisms, appraisal and governance age in years, tool specifics in months, and a curriculum that labels which is which stays honest without constant rewrites.
Should the curriculum address AI in French-language programmes distinctly?
Yes, in one specific way: verification sources and clinical-nuance translation deserve francophone worked examples, since the hazard profile differs when tools default to English-dominant training data and the assessment runs in French.