In medical education, the main question is no longer whether future physicians should learn about AI. The harder question is whether a medical school can point to a place in the curriculum where AI literacy is taught, practiced, supervised, and assessed without asking each clerkship or course director to invent the plan from scratch.
That distinction matters because the competency literature has become much more orderly than the curriculum reality. Hunt and colleagues’ 2026 scoping review synthesized 54 studies from 22 countries and reduced 564 proposed competency statements into 7 domains, 37 competencies, and 170 learning objectives for medical education.[1] A separate Delphi-consensus set organized expectations into 23 competencies: 8 knowledge competencies, 9 skills competencies, and 6 attitudes.[2] These are not casual lists of desirable topics. They are attempts to describe what a physician should be able to know, do, and judge when AI enters clinical work.
Implementation has not kept pace with that conceptual maturity. In the 2023-2024 AAMC SCOPE survey, 77% of U.S. and Canadian MD- and DO-granting schools reported covering AI, but an EDUCAUSE 2024 figure cited by AAMC found that only 14% of institutions had developed formal generative AI curricula.[3] Those two numbers should not be treated as a clean before-and-after measure. They come from different surveys, with different populations and definitions. Still, they put pressure on the same weak point: mentioning AI somewhere in a program is not the same as building a formal, assessable curriculum.
AI competency is becoming specific enough to teach
The most useful feature of the Hunt synthesis is not simply its size, although the scale is notable. Its value is that it separates a broad aspiration into curricular units that a faculty team can actually inspect. The seven-domain taxonomy covers foundational AI concepts, clinical use, evaluation and limitations, ethics and equity, data privacy and governance, communication, and professional responsibility.[1] That does not solve sequencing, staffing, assessment, or faculty development. It does give curriculum committees a map detailed enough to expose omissions.
The Delphi-consensus set is useful for a different reason. By grouping competencies into knowledge, skills, and attitudes, it makes the educational task harder to evade.[2] A student may know that a model can produce biased or unreliable output, but that is not the same as being able to question the use of an AI recommendation in a patient encounter. A student may be comfortable prompting a tool, but that is not the same as accepting professional accountability for decisions made with algorithmic support.
| Framework | What it contributes | What curriculum designers still have to decide |
|---|---|---|
| Hunt et al. seven-domain taxonomy | A broad synthesis of proposed AI competencies across 54 studies, organized into 7 domains, 37 competencies, and 170 learning objectives | Where the learning objectives belong, which are required for graduation, and how they are assessed |
| Delphi-consensus 23 competencies | A knowledge-skills-attitudes structure that is easier to translate into educational planning | How to protect attitudes and professional judgment from being treated as optional discussion topics |
| AAMC responsible AI principles | Institutional principles for transparency, equity, accountability, privacy, faculty development, student engagement, and continuous improvement | How principles become course-level expectations for learners and faculty |
| AMA November 2025 policy | A signal that model learning objectives and curricular toolkits are moving onto the national agenda | How much authority those tools will carry before accreditors issue binding expectations |
The overlap across these models is more important than their differences. They converge on a physician who understands basic AI terminology, recognizes common clinical uses, can evaluate limitations, attends to bias and equity, protects patient data, explains AI-supported care to patients, and preserves professional judgment. Those expectations now appear stable enough that a school waiting for perfect consensus risks waiting past the point of educational responsibility.
The seven domains are not seven equal lectures
A competency framework can look deceptively neat on paper. In practice, its domains do not enter the curriculum with equal weight, equal timing, or equal faculty ownership. Foundational AI literacy can sit near biostatistics, informatics, or evidence-based medicine. Clinical use belongs closer to diagnosis, therapeutics, workflow, and clerkship supervision. Evaluation of limitations requires students to ask what a model was trained on, what it was validated against, and what happens when it is used outside the setting where it performed well.
Ethics and equity are especially vulnerable to curricular underplacement. They are easy to name in a competency table and easy to lose when the implementation plan becomes a search for available lecture time. The same is true of patient communication. If students are expected to explain the role and limits of AI-supported recommendations, that expectation cannot live only in a policy statement. It has to appear somewhere learners are observed communicating, corrected when they overstate certainty, and evaluated on whether they preserve patient trust.
This is where the Delphi structure adds discipline. Knowledge competencies ask whether students understand concepts. Skills competencies ask whether they can apply, evaluate, or communicate in context. Attitude competencies ask whether they approach AI with appropriate skepticism, accountability, humility, and attention to patient welfare.[2] If a program only builds the first category, it may produce students who can define terms but cannot handle the professional ambiguity of AI-supported care.
Workforce urgency is real, but it does not write the curriculum
The clinical environment is moving quickly enough to justify curricular attention without grand claims about every future physician becoming an AI specialist. AAMC has cited more than 1,200 FDA-cleared AI algorithms as of August 2025, a figure that is sufficient to show why students cannot be trained as if algorithmic tools are peripheral to medicine.[3] For readers following the evidence side of device clearance, ClinicalMind’s discussion of the evidence gap in FDA-cleared AI medical devices is a useful companion to the education question.
But workforce urgency does not determine curricular architecture. It does not tell a school whether AI belongs in a preclerkship thread, a required clerkship activity, a capstone, a longitudinal assessment, or all of those in different forms. It does not identify which faculty are prepared to teach it. It does not settle whether students should be assessed on conceptual understanding, tool use, critique of validation evidence, patient communication, or professional accountability. The frameworks make those decisions more visible; they do not make them automatic.
Adoption data show a curriculum gap, not a lack of interest
The 77% SCOPE figure is encouraging only if it is read carefully. It suggests that most responding U.S. and Canadian medical schools were at least covering AI in some form during 2023-2024.[3] That is a meaningful shift from treating AI as an extracurricular concern. It does not say that most schools have required, sequenced, competency-based AI curricula. It does not say that students are assessed on the competencies appearing in the newer frameworks. It does not tell us whether AI appears as a single lecture, an elective, a faculty-developed module, or a longitudinal expectation.
The 14% formal generative AI curriculum figure points in the other direction, but it also needs restraint.[3] Generative AI is not identical to all clinical AI, and formal curricula are not the only place learning occurs. Students may be experimenting with tools, faculty may be adapting assignments, and clerkship teams may be creating local guidance that never appears in a formal curriculum inventory. Still, for curriculum governance, informal activity is a fragile substitute. It varies by site, depends on individual faculty confidence, and is hard to assess consistently.
The two statistics therefore do not prove hypocrisy or failure. They show two different layers of adoption. The first layer is topic visibility: AI has entered the conversation at many schools. The second layer is curricular integration: AI has been translated into formal learning objectives, teaching responsibilities, learner activities, and assessment. The gap between those layers is exactly where curriculum designers are now working.
Principles help institutions, but competencies help courses
The AAMC’s July 2025 Principles for the Responsible Use of AI in and for Medical Education occupy a different level from the competency frameworks. Version 2.0 identifies seven principles: transparency, equity, accountability, privacy, faculty development, student engagement, and continuous improvement.[4] These principles are institutionally important because they frame the conditions under which AI should be used in medical education, not only taught as content.
That makes them necessary but insufficient for course planning. A principle such as transparency can guide policy on AI disclosure in assignments or educational tools. It does not by itself tell a clerkship director what a student must demonstrate during a patient presentation when an AI-assisted differential diagnosis has been used. A privacy principle can support institutional guardrails. It does not automatically become a student-level competency unless someone writes the objective, teaches the behavior, and decides how performance will be judged.
The Macy Foundation domains, as summarized in the available synthesis, widen the frame further by considering admissions, classroom learning, workplace learning, assessment, and program evaluation.[2] That is useful because AI is not only subject matter. It may affect how applicants are reviewed, how students learn, how clinical performance is observed, and how programs evaluate themselves. For the narrower question of graduating physician competence, however, domain-level governance still has to be translated into learner-facing expectations.
Why mature frameworks still lag in classrooms
The barriers are not mysterious. A 2025 Twelve Tips paper identifies familiar obstacles, including lack of faculty expertise, curriculum crowding, and the need for phased implementation.[5] These are not minor operational complaints. They determine whether a competency framework is usable by the people who inherit it.
Faculty expertise is the first bottleneck because AI competencies cross several knowledge communities. A data scientist may explain model performance, a clinician may explain workflow consequences, an ethicist may explain bias and accountability, and a communication faculty member may observe how a student discusses uncertainty with a patient. A curriculum plan that assumes one faculty group can cover the whole field will usually either narrow the content too much or leave the professional dimensions underdeveloped.
Curriculum crowding is the second bottleneck because AI rarely arrives with unused time attached. Adding it as a freestanding requirement may be unrealistic in some programs. Embedding it into evidence-based medicine, clinical reasoning, informatics, ethics, radiology, pathology, or clerkship teaching may be more plausible, but embedded content is also easier to make invisible. If it is not mapped and assessed, it can disappear into a course description while students receive uneven exposure.
Assessment is the third bottleneck, and it is where the difference between a topic and a competency becomes visible. If students are expected to evaluate an AI-supported recommendation, a school has to decide what counts as adequate evaluation. If they are expected to communicate AI limitations to patients, someone has to observe that communication. If they are expected to recognize bias, privacy risk, or automation overreliance, those expectations need more than a slide in a lecture.
The missing mandate matters, even if it is not the whole story
As of Q3 2026, no single AI competency framework has been mandated by the major U.S. medical education accreditors, including LCME for undergraduate medical education and ACGME for graduate medical education. That absence does not mean schools are unable to act. Medical education often changes before accreditation language catches up. It does mean local faculty teams are working in a crowded space of credible but nonbinding guidance.
The consequence is predictable. One institution may start with the Hunt taxonomy because it wants comprehensive mapping. Another may prefer the Delphi knowledge-skills-attitudes structure because it fits existing competency-based education processes. Another may anchor policy in AAMC responsible-use principles and then build student objectives locally. These choices are defensible, but they also make cross-school comparability difficult. A graduating student’s AI preparation may depend less on national consensus than on whether a local curriculum committee has the time, expertise, and authority to translate that consensus.
The AMA policy increases pressure for standardization
The AMA’s November 2025 policy is important because it moves the issue from professional concern toward curricular infrastructure. The policy resolves to develop model AI learning objectives and curricular toolkits across the medical education continuum.[6] That is not the same as an accreditation requirement, and it is not proof that implementation is solved. It is a signal that national organizations are beginning to supply the missing connective tissue between broad principles and local course design.
Model objectives could help schools avoid duplicating basic work. Toolkits could give faculty a starting point for teaching and assessment. But their practical effect will depend on whether they are specific enough to be adopted without extensive reinvention, flexible enough for different curricula, and aligned with the competency domains already emerging from the literature. If they become another polished layer of guidance without clear links to courses and assessments, they will add legitimacy without reducing much workload.
What curriculum designers can reasonably conclude in 2026
Medical schools now have enough evidence-based competency material to begin alignment. The Hunt synthesis provides breadth and granularity. The Delphi set provides an educational structure that protects skills and attitudes from being collapsed into knowledge. The AAMC principles provide institutional guardrails. The AMA policy suggests that model objectives and toolkits are on the way. None of these, alone, supplies a binding national standard.
The immediate risk is not that schools lack any framework to choose from. The risk is that framework maturity will be mistaken for institutional readiness. A competency can be well described in the literature and still have no assigned course, no prepared faculty, no learner activity, and no assessment. Conversely, a school can report that it covers AI while leaving students with uneven exposure and no clear expectation for clinical use, critique, communication, or accountability.
By Q3 2026, the decision environment is clearer but not settled: multiple credible frameworks, partial adoption, no LCME or ACGME mandate, growing clinical exposure to AI tools, and increasing pressure from professional organizations. The next curriculum question is no longer whether AI belongs in medical education. It is which competencies a school is willing to own, teach, and assess.
References
- What the AI era doctor should know: a scoping review of proposed artificial intelligence competencies for medical education, npj Digital Medicine, 2026.
- Future of Medical Education, Physician AI Handbook.
- Medical schools move from worrying about AI to teaching it, AAMC News.
- Principles for the Responsible Use of AI in and for Medical Education, AAMC, July 2025.
- Twelve tips for integrating artificial intelligence into medical education, Medical Education Online, 2025.
- AMA adopts policy to advance AI literacy in medical education, American Medical Association, November 2025.
Comments
Join the discussion with an anonymous comment.