The conversation about artificial intelligence in health professions education often begins with competencies. What should students know about AI? How should lecturers use it? Which tools should a university approve?
Those questions matter, but they do not tell us what conditions allow professional learning to develop when AI is present. A learner can use an approved tool, disclose it correctly and produce an excellent answer while still bypassing the dialogue, uncertainty and reflection through which professional judgement develops.
Michael Rowe's theoretical framework for integrating AI into health professions education offers a useful change of direction. Instead of starting with the tool, it starts with how professional learning works.
The problem with focusing only on capability
Many AI frameworks describe what people need to know: literacy, competencies, ethical awareness and appropriate tool use. Rowe argues that they leave another question open: what conditions support professional learning with AI?
Without a pedagogical answer, integration can become defensive, treating AI mainly as a threat to control, or opportunistic, pursuing efficiency while the quality of learning quietly erodes. If AI can produce a plausible clinical-reasoning chain or reflective entry, we need to ask what the task was actually designed to develop.
Four theories, one design challenge
The framework maps four perspectives across learning interactions. Social constructivism asks how knowledge is constructed through dialogue. Critical pedagogy asks whose assumptions and interests shape what counts as knowledge. Complexity theory locates practice in dynamic, relational and context-dependent systems. Connectivism treats knowledge as distributed across people, disciplines, tools and networks.
Six areas of convergence become design principles. They apply whether or not AI is present, but AI makes each one more urgent because it can either support the learning condition or allow the learner to bypass it.
1. Dialogic knowledge construction
Knowledge is constructed through dialogue rather than simply transferred. When AI generates content for a learner to accept, even an accurate response can produce surface understanding.
The design response is to make learners evaluate, challenge, compare or extend the output. The dialogue is the learning—not merely the answer it produces.
2. Critical consciousness
Learners need to recognise what shapes accepted knowledge, including an AI system's assumptions, omissions, limitations and embedded interests.
In a clinical case, students might compare AI reasoning with their own reasoning and ask what each surfaces, what each conceals and why the difference matters for practice and for the patient.
3. Adaptive expertise
Professional practice is not the reproduction of a known procedure. It requires action when the procedure does not fit. Reproductive answers are increasingly easy for AI to generate; responding safely to contradictions, edge cases and changing information is not.
Learning activities should therefore require adaptation across unfamiliar cases rather than reward the reproduction of a standard plan.
4. Contextual authenticity
Healthcare is complex, relational and particular. Removing social determinants, resource constraints, competing priorities and communication difficulty may make a case easier to mark while making it less useful preparation for practice.
AI can help educators introduce meaningful variation and complexity. It should not be used to simplify away the very conditions that make professional judgement necessary.
5. Metacognitive development
A correct output can create the feeling that learning has occurred. Without awareness of their own reasoning, learners may struggle to distinguish genuine understanding from its appearance.
Students should narrate their initial judgement, uncertainty, use of AI and subsequent revisions. This makes the reasoning process visible and assessable rather than leaving only a polished conclusion.
6. Networked knowledge building
Healthcare problems cross professional and disciplinary boundaries. Learning that remains within one silo can leave graduates unable to navigate the knowledge networks that complex practice demands.
AI may help surface perspectives from other disciplines, but learners still need to decide which expertise is relevant, how perspectives relate and where collaboration with a real person is required.
Design constraints, not another checklist
The six principles are most useful before an activity is built. For each proposed AI interaction, I would ask: which learning condition is this intended to support, how could the same interaction undermine it, and what learner action would make the intended process visible?
At curriculum and assessment level, the framework invites another difficult question: are we measuring production of an artefact, or the reasoning behind it? AI may not have created that weakness. It may simply have made it impossible to ignore.
How this complements other frameworks
The decision guide ‘Should I Use AI for This?’ asks whether an individual should use AI for a task. Health CARE-AI asks whether that use is contextual, accountable, responsible and equitable. Rowe's framework adds a third question: Does the design preserve the conditions under which professional learning develops?
All three questions are needed. An AI activity can be personally defensible and ethically governed yet remain educationally weak if it bypasses dialogue, adaptation, context or metacognition.
A theoretical proposal, not proof of effectiveness
This framework is a theoretical synthesis presented as a preprint. The six principles were derived through conceptual mapping across learning theories, not through a trial showing improved learner or patient outcomes.
That makes the principles propositions to design with and study, not settled rules. Research now needs to translate them into observable educational practices, examine how learners and educators experience them, and test whether they improve reasoning, adaptation, transfer or metacognitive accuracy without creating new inequities or workload.
The question that should stay constant
AI capabilities will continue to change. The pedagogical question does not need to change with every new model: Are we creating the conditions in which learners develop the reasoning, judgement and adaptive capacity that healthcare demands?
If an AI activity cannot answer that question, its novelty or efficiency is not enough. The aim is not merely to help students produce better work. It is to design the kind of participation through which they become better professionals.
Reference
- Rowe, M. (2025). A theoretical framework for integrating AI into health professions education [Preprint]. Open Science Framework. https://doi.org/10.31219/osf.io/c764f_v2.