The adoption statistic is genuinely striking: the proportion of US and Canadian medical schools incorporating AI into their curricula rose from 53% in 2023 to 77% in 2024, per AAMC-reported figures, one of the fastest curricular shifts in modern medical education. The number deserves both its headline and its interrogation, because "incorporating AI" spans everything from a single optional lecture to an assessed longitudinal competency, and the difference between those is the difference between a curriculum and a checkbox. This article builds the hierarchy that makes the 77% readable, and asks the question the statistic cannot answer: what are students actually learning?
The integration hierarchy
Six levels, ascending. Optional exposure: a guest lecture or elective session; AI exists, attendance optional, assessment none. Elective depth: a selected-students course, real content for the self-selected few. Embedded exercises: AI woven into existing teaching, an AI-assisted differential exercise inside a clinical block, reaching everyone at low intensity. Assessed competency: students are examined on AI-related knowledge or skill, the level at which incorporation acquires teeth, because assessment drives learning here as everywhere. Clinical-placement use: supervised, governed use of AI tools in real clinical learning environments, where the literacy meets its context. Longitudinal curriculum: a planned thread across years, building from mechanisms to appraisal to supervised application, the level the 77% headline implies and the minority of programmes likely occupy. A school at level one and a school at level six both count in the statistic; students comparing programmes should ask which level, and the question itself signals what to look for.
What should the content be?
The substantive question underneath the levels. The plausible curriculum, aligned with the direction of the AAMC's competency-development work, spans: how generative systems work at working-concept level, plausibility engines, grounding, stylistic confidence; model appraisal, reading claims by outcome rung, /blog/ai-learning-outcome-ladder-medical-education, and knowing the failure modes by feel; ethics and governance, bias, equity, accountability; confidentiality in practice, what never enters a consumer tool; documentation and disclosure norms; and supervision skill, the ability to check AI work rather than merely use AI tools, which is the capability clinical practice will actually demand. Prompt technique appears in this list nowhere near the top, deliberately: prompting is the perishable layer, appraisal and supervision are the durable ones, and a curriculum weighted toward the former has taught this year's interfaces rather than a career's judgement.
Reading the gap between adoption and use
The adoption statistic sits beside a student-behaviour statistic that reframes it: Canadian survey data show 78.9% of responding medical students already using generative AI, 53% at least weekly, and 75.9% wanting formal incorporation, with strong majorities simultaneously recognising inaccuracy and bias risks. Students, in other words, arrived before the curricula did, and much current incorporation is institutions catching up to established behaviour rather than introducing a novelty. That ordering matters for design: teaching that treats students as AI-naive misses its audience, while teaching that starts from real student workflows, then adds the appraisal, governance and supervision layers students cannot self-teach, meets the actual gap. The same pattern, high informal use, uncertain formal outcomes, is the global story of this transition, and the curricula that close it will be the ones that assess what they claim to teach.
Frequently asked questions
Is 77% good news?
Directionally yes, adoption at speed signals institutional seriousness; the hierarchy is the caution, because the statistic counts presence, not depth, and depth is what changes practice.
What should an applicant or student ask a school?
Three questions: at which level of the hierarchy does AI sit here; is any of it assessed; and does it reach clinical placements with governance, the answers locate a programme faster than any brochure.
How does this compare with the UK and Australia?
Different frameworks, converging cores; the four-country comparison, MSC and HDR UK recommendations, AAMC competencies, the AMC framework and Canada's ecosystem, is at /blog/medical-ai-literacy-uk-us-canada-australia.
Will AI content appear in USMLE-style examinations?
Assessment bodies publish their own blueprints and any change lands there first; the safer prediction is indirect, AI literacy shaping how schools teach and assess locally before national examinations move, which is one more reason the hierarchy's assessment level matters.
What can students at low-hierarchy schools do meanwhile?
Self-build the durable layers: appraisal by outcome rung, verification workflow, disclosure habits and supervised-use judgement; every one is learnable without institutional permission, and the frameworks converging internationally describe the syllabus for free.
Is there a risk of overcorrection, too much AI content in curricula?
A real one worth naming: curriculum hours are zero-sum, and AI literacy should displace redundancy, not clinical fundamentals; the six durable capabilities are compact by design, and bloated versions usually signal prompting workshops padding out the timetable.