A theoretical framework for integrating AI into health professions education

About this essay

• Author: Michael Rowe (ORCID)
• Affiliation: University of Lincoln (mrowe@lincoln.ac.uk)
• Created: March 29, 2025
• Version: 1.0 (last updated: April 5, 2026)
• Keywords: artificial intelligence, complexity, connectivism, critical pedagogy, health professions education, learning theory, social constructivism
• License: Creative Commons Attribution 4.0 International
• Preprint DOI:
  • Open Science Framework (preferred): 10.31219/osf.io/c764f_v1
  • Zenodo (legacy): 10.5281/zenodo.15106214
• Peer reviewed: No

Abstract

Health professions education has a problem that predates AI but that AI has made impossible to ignore: our educational approaches have tended to prioritise content transmission and artefact production over the conditions under which professional learning actually occurs. Graduates emerge simultaneously overwhelmed with information and underprepared for complex practice. Students are already using AI extensively — their motivations for doing so are inevitably mixed, but the pattern itself suggests that existing educational structures leave needs unmet that these tools address.

This paper develops a theoretically grounded set of design principles for integrating AI into health professions education (HPE). Drawing on social constructivism, critical pedagogy, complexity theory, and connectivism, I conducted a structured conceptual analysis of learning interactions across six dimensions: power dynamics, knowledge representation, agency, contextual influence, identity formation, and temporality. Six convergences emerged — dialogic knowledge construction, critical consciousness, adaptive expertise, contextual authenticity, metacognitive development, and networked knowledge building — each identifying conditions under which learning is effective according to multiple theoretical perspectives. These convergences were translated into design principles grounded in how learning works rather than in current AI capabilities. The principles would support effective education with or without AI. They shift attention from policing outputs to creating the conditions under which students develop the reasoning, judgement, and adaptive capacity that healthcare demands.

Introduction

Health professions education produces a troubling tension: graduates who are simultaneously overwhelmed with information and underprepared for the complexity of modern practice (Frenk et al., 2010; Irby et al., 2010). Content overload, decontextualised knowledge, and assessment practices that reward recall over reasoning — these are not new problems. They are well-documented features of educational approaches increasingly at odds with what healthcare delivery actually requires (Frenk et al., 2010; Van Der Vleuten, 2016).

The challenges are entrenched. Students struggle to integrate theoretical knowledge with clinical practice. Educators labour under curricular structures that fragment learning into discrete subjects and skills. Assessment practices measure what is easy to evaluate rather than what is important to learn — treating outputs as proxies for learning rather than observing the processes of learning itself. The gap between how we educate and how professionals actually work continues to widen, despite sustained efforts to bridge it.

Artificial intelligence has entered this landscape not as a solution, nor merely as another technology to integrate, but as something more disruptive: a catalyst that exposes assumptions we had stopped examining. Students are already using AI tools extensively — surveys consistently show adoption rates of 80–90% — to explore concepts, practise reasoning, and prepare for assessments, regardless of whether educators have sanctioned or acknowledged this use. Some of this use is straightforwardly instrumental: students looking for shortcuts, just as they always have. But the scale and consistency of adoption suggests something beyond convenience-seeking. The interesting question is not how to control this adoption but what students’ rapid embrace of AI reveals about the limitations of what we were already doing.

The answer is uncomfortable. AI does not introduce new problems into health professions education so much as it makes existing ones harder to ignore. When students use AI to generate assignment text, this reveals that our assessments were measuring the difficulty of producing artefacts rather than the depth of learning behind them (Dawson et al., 2024). When AI can summarise a lecture more efficiently than attending it, this exposes the poverty of transmission-based teaching. When AI support improves students’ short-term task performance while simultaneously reducing their metacognitive engagement — a pattern Fan et al. (2024) describe as “metacognitive laziness” — this illustrates how AI can produce the appearance of learning while undermining its substance. The challenge is not to defend existing practices against AI but to recognise what AI’s disruptive presence tells us about those practices — and to respond with approaches grounded in how learning actually works.

Several frameworks address aspects of AI in education. TPACK (Mishra & Koehler, 2006) examines the intersection of technological, pedagogical, and content knowledge. UNESCO’s AI competency framework (2023) outlines capabilities for learners and educators. Ng et al.’s (2021) AI literacy framework identifies competencies for understanding and using AI, extended by Chee et al. (2025) into a comprehensive competency framework mapping AI literacy across different learner groups and educational stages. These are useful, but they share a common orientation: they focus on what people need to know about AI. The question they leave open is how established learning theory should shape the integration of AI into educational practice.

Recent work has begun to address this gap. Wegerif and Casebourne (2025) propose a “double dialogic pedagogy” grounded in dialogic theory, arguing that AI should be positioned as a partner in expanding the space of dialogue rather than as a replacement for human thinking. Their contribution is valuable — particularly in articulating how AI might support both immediate dialogic exchange and longer-term induction into cultural conversations. However, their framework draws on a single theoretical tradition (dialogic theory), and their focus is general education rather than the specific demands of professional formation in complex, high-stakes practice environments. The present paper complements and extends this work by integrating multiple theoretical perspectives and grounding the analysis in the particular challenges of health professions education — a context characterised by the combination of high-stakes clinical practice, rapidly evolving knowledge, acute tension between standardised assessment and adaptive expertise, and the formative challenge of developing professional identity under conditions of genuine uncertainty. HPE does not merely illustrate general educational challenges; it concentrates them in ways that make the stakes of getting AI integration right — or wrong — particularly visible.

This paper starts from theoretical foundations about the nature of learning and works forward to design principles for AI integration. I examine four interrelated dimensions of learning through complementary theoretical lenses: the how of learning (social constructivism), the why (critical pedagogy), the where (complexity theory), and the what (connectivism). The aim is to move past the framing of AI as either threat or panacea and toward a more productive question: under what conditions does professional learning occur, and how can AI be integrated in ways that support rather than undermine those conditions?

Theoretical foundations

The analysis draws on four theoretical perspectives, each addressing a dimension of learning the others do not fully capture. Other theories could have been included. Self-determination theory (Deci & Ryan, 2000) would foreground motivation and autonomy. Activity theory (Engeström, 2001) would emphasise mediated action within systems. Phenomenological approaches (Dall’Alba & Barnacle, 2007) would centre embodied experience. I chose these four not because they exhaust the landscape, but because their complementarity across the how, why, where, and what of learning provides analytical breadth without redundancy. The positionality statement below provides additional context for this selection.

Social constructivism

Social constructivism addresses the how of learning. Knowledge, in this view, is not transferred from teacher to student but actively constructed through social interaction (Vygotsky, 1978). This is a familiar claim, but its implications are routinely ignored in HPE, where content delivery and procedural training remain dominant. Vygotsky’s zone of proximal development — the space between independent capability and what becomes possible with guidance — describes something clinical educators recognise intuitively: students develop competence through scaffolded interaction with experienced practitioners, not through information transfer (Jonassen, 1995). If knowledge is constructed through social processes, AI should serve as a dialogical partner supporting meaning-making, not as a content generator producing outputs that students passively consume.

Critical pedagogy

Critical pedagogy addresses the why of learning — what education is for. Freire’s (2000) critique of “banking education,” where educators deposit knowledge into passive students, describes something still pervasive in HPE: curricula that prioritise standardised content delivery over student agency, reinforcing the authority of institutions while positioning students as recipients of established knowledge (hooks, 1994). The alternative Freire proposed — conscientisation, the development of critical awareness of the systems shaping one’s experience — is directly relevant to AI integration. The question is not simply whether students can use AI effectively. It is whether they can evaluate whose interests are served by the knowledge AI produces, what assumptions are embedded in its outputs, and what is rendered invisible by its design. AI interactions that develop this critical consciousness are educationally valuable. AI interactions that bypass it are not.

Complexity theory

Complexity theory addresses the where of learning — the systems in which education and healthcare practice occur. HPE has been structured around linear, reductionist models: modular curricula, sequential skill development, standardised procedures, predictable outcomes (Fraser & Greenhalgh, 2001). Contemporary healthcare operates differently. It is a complex adaptive system characterised by nonlinearity, emergence, and unpredictability (Plsek & Greenhalgh, 2001; Bleakley, 2010). The misalignment is fundamental. Educational approaches that emphasise memorisation of procedures and discrete skill development produce practitioners who are underprepared for environments where they must constantly adapt knowledge to unique presentations and evolving conditions. AI’s contribution here is not to deliver content more efficiently but to create dynamic learning experiences that mirror the complexity of actual practice.

Connectivism

Connectivism addresses the what of learning — the nature and organisation of knowledge. It starts from the observation that knowledge no longer resides primarily in individuals or static repositories but is distributed across networks (Siemens, 2005; Goldie, 2016). What students need to learn shifts accordingly: from static content to the connections between knowledge domains, the patterns linking disparate areas, and the capacity to navigate evolving knowledge networks. This reframing matters because healthcare specialisation continues to deepen while the problems practitioners face increasingly require knowledge that crosses disciplinary boundaries. AI can make visible connections between knowledge domains that might otherwise remain obscure — but only if it is used to support network navigation rather than content delivery.

Analytical approach

Methodology

Moving from theoretical foundations to design principles requires an approach that maintains theoretical integrity while producing something practically useful. I adopted a structured conceptual analysis (Jabareen, 2009) involving three phases: constructing an analytical lens through which to compare the theories, mapping each theory’s propositions across that lens, and identifying convergences — points where multiple theories arrive at similar insights despite different conceptual starting points.

Constructing the analytical lens

To compare the four theories systematically, I needed a consistent frame. I identified six dimensions of learning interactions — the points of engagement between learners and other elements of the educational ecosystem (Chi, 2009; Stahl et al., 2006). These dimensions emerged from close reading of the four traditions, representing the fundamental concerns each theory addresses, albeit differently: power dynamics (how authority is distributed), knowledge representation (how knowledge is structured and engaged with), agency (how learner autonomy is supported or constrained), contextual influence (how broader contexts shape learning), identity formation (how professional identity develops), and temporality (how learning unfolds over time).

These are not arbitrary categories. They are consistent with established analytical frameworks in learning theory. Biggs’s (1993) 3P model addresses factors overlapping several of these dimensions. Illeris (2003) identifies content, incentive, and interaction as the three dimensions of learning. Biesta (2015) distinguishes qualification, socialisation, and subjectification as functions of education — mapping onto the knowledge, identity, and agency dimensions. Barnett (2000) emphasises uncertainty and complexity as conditions of higher education, captured across the contextual and temporality dimensions. I could have included additional dimensions — most notably embodiment (Dall’Alba & Barnacle, 2007) and affect — but these are less consistently theorised across all four perspectives, making systematic comparison more difficult. Their omission is a limitation I return to later.

Positionality

I should be direct about what I brought to this analysis. Fifteen years of teaching and researching the use of technology in professional education has left me with commitments about what effective technology-enhanced learning looks like. In qualitative research terms, this is theoretical sensitivity (Glaser & Strauss, 1967) — the capacity to recognise meaningful patterns because you have deep domain knowledge. My selection of these four theories reflects not only their complementary coverage but a longstanding intellectual engagement with constructivist, critical, and complexity-informed approaches to education. The convergences I identified were shaped by this engagement.

I have tried to maintain rigour by grounding claims in the theoretical texts and making the comparative process transparent. But I am not claiming to have discovered the objective structure of learning theory. The resulting principles represent the intersection of theoretical convergence and informed professional judgement — one map through a problem space. Other scholars with different commitments would construct different maps. This is a feature of conceptual analysis, not a limitation to be overcome.

Comparative analysis

I analysed each theory’s propositions across the six dimensions, creating a comparative matrix (Table 1) to make convergences and tensions visible (Suthers & Rosen, 2011; Wong et al., 2013).

Table 1. Comparative matrix: four theories across six dimensions of learning interactions

Dimension	Social constructivism	Critical pedagogy	Complexity theory	Connectivism
Power dynamics	Gradual transfer of authority from expert to learner through fading scaffolding	Explicit critique of power hierarchies; redistribution through dialogue	Challenges hierarchical control; emphasises self-organisation and distributed decision-making	Authority distributed across networks; fluid and context-dependent
Knowledge representation	Personally meaningful constructions; provisional, evolving through collaborative meaning-making	Never neutral; reflects interests and power structures; values experiential knowledge	Emergent patterns rather than fixed facts; provisional, contextual, evolving	Exists in connections between sources; distributed, accessed not possessed
Agency	Learners as active constructors; autonomous exploration within scaffolded boundaries	Conscientisation; learners as subjects who act upon their world	Distributed across the system; adaptive responses rather than predetermined actions	Capacity to build, navigate, and reconfigure networks; decision-making about what to learn
Contextual influence	Inseparable from social and cultural context; authentic contexts essential	Embedded in sociopolitical contexts; addresses real-world problems	Context constitutive of learning; interdependent systems	Networks extend beyond formal contexts; quality depends on diversity
Identity formation	Becoming a community member through legitimate peripheral participation	Recognising one’s position within social systems; developing capacity to transform them	Emerges from participation in complex adaptive systems; comfort with uncertainty	Develops through participation in knowledge networks; practitioners as nodes
Temporality	Developmental trajectories; progressive participation; timing of support crucial	Cycles of action and reflection (praxis); ongoing conscientisation	Non-linear progression; stability punctuated by rapid change	Real-time engagement with evolving networks; compressed learning cycles

Six convergences

What matters about this matrix is not what each theory says individually but where they converge despite their different starting points. Six convergences emerged — conditions under which learning is effective according to multiple theoretical perspectives.

1. Dialogic knowledge construction. None of these theories treats knowledge as something that can be transmitted. The mechanisms differ — scaffolded meaning-making in constructivism, liberatory exchange in critical pedagogy, emergent understanding from system interactions in complexity theory, knowledge created through network connections in connectivism — but the convergence is clear: learning that bypasses dialogue produces shallow understanding regardless of the quality of the output. This has obvious implications for AI integration, where the temptation is to use AI precisely to bypass dialogue in favour of efficient content generation.

2. Critical consciousness. Conscientisation is critical pedagogy’s term, but the imperative to evaluate the conditions shaping knowledge runs through all four perspectives. Constructivism demands critical reflection on community assumptions. Complexity theory requires awareness of how systemic dynamics constrain and enable action. Connectivism requires evaluation of information sources, network quality, and — increasingly — algorithmic mediation. The convergence: effective learning requires awareness of the forces that shape what counts as knowledge. Without this, students consume AI outputs the same way they consume textbooks in banking education — uncritically.

3. Adaptive expertise. The question is not whether students can reproduce known procedures but whether they can act effectively when procedures do not apply. All four theories challenge reproductive expertise: constructivism through progressive scaffolding toward independence, critical pedagogy through the capacity to transform rather than merely respond, complexity theory through adaptive responses to nonlinear conditions, connectivism through navigation of evolving knowledge networks. This distinction — adaptive versus reproductive expertise — becomes particularly salient when reproductive outputs are computationally trivial to generate.

4. Contextual authenticity. Every theory insists that separating learning from context impoverishes it. Constructivism stresses situated learning. Critical pedagogy demands engagement with real conditions. Complexity theory treats context as constitutive of learning, not merely a backdrop. Connectivism locates learning in diverse network contexts. Decontextualised knowledge — the kind that can be delivered efficiently through lectures and assessed through standardised examinations — misses what matters most about professional practice: that it happens in conditions of complexity, uncertainty, and human particularity. This convergence also carries implications for curriculum design: if contexts are dynamic, curricula that respond to evolving conditions serve learners better than rigid, predetermined structures.

5. Metacognitive development. Each theory emphasises awareness of one’s own thinking, framed differently: self-regulation in constructivism, praxis in critical pedagogy, pattern recognition in complexity theory, meta-learning in connectivism. The convergence matters because without metacognitive awareness, students cannot distinguish between recognising a correct output and understanding the reasoning behind it. This is not a theoretical abstraction — Fan et al. (2024) demonstrated empirically that students using AI showed improved short-term task performance while exhibiting significantly fewer metacognitive processes, a pattern the authors describe as “metacognitive laziness.” The distinction between using AI to produce an answer and using AI to develop judgement is the one that most institutional responses to AI have not yet adequately addressed.

6. Networked knowledge building. Learning that stays within disciplinary silos impoverishes the knowledge available to practitioners. Constructivism’s communities of practice can extend beyond professional boundaries. Critical pedagogy challenges disciplinary silos as power structures constraining whose knowledge counts. Complexity theory emphasises emergent knowledge from interconnections across system components. Connectivism explicitly positions learning as the creation and navigation of diverse networks. The convergence: knowledge building is strengthened by connections across boundaries of discipline, institution, and epistemology — a condition that AI is uniquely positioned to support through its capacity to surface connections across distributed knowledge sources.

These convergences describe conditions for effective learning. They are not contingent on current AI capabilities or failure modes. They would support effective education with or without AI. But the introduction of AI makes them newly urgent, because AI can either support these conditions or undermine them depending entirely on how it is integrated.

Design principles for AI integration

The move from descriptive convergence to prescriptive principle involves an interpretive step that should be made explicit. The logic is straightforward: if multiple learning theories converge on a condition under which learning is effective, then AI integration that supports this condition is likely to enhance learning, while integration that contradicts it is likely to undermine learning. The six principles below apply this reasoning to each convergence. Table 2 illustrates what each looks like in practice — both with and without AI — because the principles describe good education, not just good AI use.

The six principles:

1. Dialogic knowledge construction. Position AI as a participant in dialogic learning rather than an authoritative source, augmenting the social processes through which knowledge is constructed.

2. Critical consciousness. Use AI to develop critical evaluation — of AI outputs, of the systems producing them, and of the power dynamics embedded in both — rather than uncritical consumption of generated content.

3. Adaptive expertise. Use AI to develop flexible knowledge application in novel situations, not efficient reproduction of standardised responses.

4. Contextual authenticity. Use AI to enhance the authentic complexity of learning contexts — incorporating social, systemic, and relational factors — rather than abstracting from that complexity.

5. Metacognitive development. Use AI to make thinking visible, surface cognitive patterns, and develop students’ capacity to monitor their own reasoning and biases.

6. Networked knowledge building. Use AI to facilitate knowledge building across disciplinary, institutional, and epistemological boundaries, supporting diverse knowledge networks rather than reinforcing silos.

Table 2. Principles in practice: with and without AI

Principle	Without AI	With AI
Dialogic knowledge construction	Case-based group discussion with structured peer challenge	AI generates alternative clinical interpretations for students to evaluate and debate
Critical consciousness	Students analyse how clinical guidelines reflect particular populations and values	Students compare AI diagnostic reasoning with clinician reasoning, examining assumptions in both
Adaptive expertise	Progressive case sequences with increasing complexity and ambiguity	AI generates adaptive scenarios responding to demonstrated competence, introducing context-specific complications
Contextual authenticity	Placement learning in diverse clinical settings	AI creates simulations incorporating social determinants, resource constraints, and communication challenges
Metacognitive development	Reflective journals analysing clinical reasoning over time	AI tracks reasoning patterns across cases, surfacing tendencies and prompting targeted reflection
Networked knowledge building	Interprofessional education with shared case analysis	AI surfaces connections across disciplinary knowledge bases, translating between professional perspectives

These principles constitute a framework in the sense that the theoretical foundations, analytical method, and resulting principles form an integrated whole. They are not grounded in what AI currently can or cannot do. They are grounded in what learning theory identifies as the conditions under which professional learning is effective. This means they do not need revision each time AI capabilities change. They are positioned neither defensively — protecting education from AI — nor opportunistically — exploiting AI for efficiency. They describe what good education looks like, and they offer guidance for ensuring that AI serves that vision rather than replacing it with something more convenient but less educationally sound.

Discussion

Relationship to existing frameworks

The distinction between this framework and existing ones is not a matter of quality but of level. TPACK (Mishra & Koehler, 2006) asks whether an educator has the knowledge to use a technology effectively. The principles developed here ask whether the way a technology is being used supports the conditions under which learning occurs. An educator can have excellent TPACK and still deploy AI in ways that undermine dialogic knowledge construction — by treating AI-generated content as authoritative rather than dialogic, for instance. Ng et al.’s (2021) AI literacy framework identifies competencies for understanding AI, but competency in using AI is not the same as using it in ways that develop adaptive expertise or critical consciousness. Chee et al.’s (2025) extension of AI literacy into a developmental pathway across educational stages provides useful granularity about what learners should know at different levels, but it remains oriented toward individual capability rather than the pedagogical conditions that shape whether those capabilities lead to genuine learning. These existing frameworks address individual capabilities. The present framework addresses the pedagogical conditions in which those capabilities are exercised.

Wegerif and Casebourne’s (2025) double dialogic pedagogy operates closer to the level of the present framework, grounding AI integration in a theory of learning rather than a theory of technology use. Their emphasis on dialogue as the mechanism through which AI becomes educationally productive aligns with the dialogic knowledge construction principle developed here. Where the present framework extends their contribution is in two directions. First, by drawing on multiple theoretical traditions rather than one, it identifies conditions for effective learning — critical consciousness, adaptive expertise, contextual authenticity, metacognitive development, networked knowledge building — that dialogic theory alone does not fully capture. Second, by grounding the analysis in health professions education, it engages with the particular demands of professional formation in complex, high-stakes environments where the relationship between what students learn and what practitioners do carries consequences for patient safety and wellbeing.

Physical access is not epistemological access

Core to the framework is a claim that most students are not natural autodidacts. The widespread availability of AI has been accompanied by an assumption — in policy, popular discourse, and sometimes educational research — that access to powerful tools naturally enhances learning. This assumption recapitulates a pattern well-documented in educational technology research. Mitra’s (2003) “hole in the wall” experiments, often cited as evidence that children teach themselves with technology, showed that unsupported access produced surface-level exploration rather than deep learning. Warschauer’s (2004) research on technology and social inclusion demonstrated that providing physical access to technology without epistemological access — the knowledge, practices, and dispositions needed to learn with it (Morrow, 2009) — consistently failed to produce anticipated outcomes.

The distinction matters for AI. A student who uses AI to generate a well-structured essay has physical access to the technology but may lack the epistemological access to learn through the process. AI can create a compelling illusion of learning: reviewing an AI-generated output and recognising it as correct produces a subjective sense of understanding while requiring none of the constructive cognitive effort that learning theories identify as essential. This is the fluency illusion (Bjork & Bjork, 2011) — re-reading material feels like learning because it produces familiarity, but familiarity is not comprehension. Emerging empirical evidence supports this concern: Fan et al. (2024) found that students using AI exhibited significantly fewer self-regulated learning processes such as evaluation, monitoring, and orientation compared to those receiving human support, even as their immediate task outputs improved. Xu et al. (2025) further demonstrated that metacognitive support is critical for maintaining effective self-regulation in AI environments — without it, students’ self-regulated learning declined. The principles address this gap by specifying conditions — dialogue, critical evaluation, adaptation, reflection — under which AI interaction produces genuine learning rather than the comfortable feeling that learning has occurred.

Networked knowledge building and wicked problems

The principle of networked knowledge building deserves particular attention because the problems it addresses are not going away. The most pressing challenges in healthcare — health inequities, the management of complex multimorbidity, service integration across institutional boundaries, the health consequences of climate change — are what Rittel and Webber (1973) called wicked problems: challenges defined by incomplete knowledge, stakeholder disagreement, and interconnection with other problems such that any intervention produces unanticipated consequences. Wicked problems cannot be addressed within disciplinary silos or through standardised procedures. They require the capacity to connect knowledge across boundaries and synthesise perspectives that may appear contradictory (Head & Alford, 2015).

AI systems are increasingly positioned as agents within knowledge networks — not merely tools for individual use but autonomous or semi-autonomous participants that surface connections across distributed knowledge sources, translate between disciplinary frameworks, and identify patterns invisible to any single human perspective. Future healthcare practitioners will work not only with AI tools but within networks where AI agents actively participate in knowledge construction. The principle of networked knowledge building prepares students for this reality: developing the capacity to build, navigate, and critically evaluate knowledge networks in which both human and artificial intelligence contribute.

Beyond individual pedagogy

The implications extend beyond what individual educators do in their classrooms. The principles challenge assessment systems that measure artefact production rather than learning processes, curriculum structures that privilege content delivery over adaptive learning, and quality assurance frameworks that reward standardisation over emergence. Institutions that treat AI integration as a matter of updating individual teaching practices while leaving these structures untouched may struggle to realise the potential these principles describe. Integrating AI into HPE is not purely a pedagogical challenge. It is an organisational one.

Tensions

The six principles are complementary but not frictionless. Critical consciousness requires scepticism toward AI outputs that could inhibit the exploratory openness that dialogic knowledge construction encourages. Adaptive expertise implies comfort with uncertainty, while metacognitive development requires structured reflection that may slow the adaptive process. Contextual authenticity demands engagement with messy practice realities that can be in tension with the scaffolding effective learning requires.

These tensions are features, not weaknesses. Learning itself involves navigating tensions — between structure and freedom, critique and engagement, individual development and collective practice. The principles are considerations to be balanced in context, requiring the same adaptive expertise and professional judgement they aim to develop in students.

Limitations

This is a conceptual analysis, not an empirical study. The principles are derived from theoretical convergences, not from observed effects on student learning; they require empirical validation in diverse HPE contexts. The analysis was conducted by a single author without inter-rater reliability or formal peer debriefing. I have been transparent about the role of prior commitments, but independent validation would strengthen the claims made here.

The four theoretical perspectives do not exhaust the relevant landscape. Other selections — including theories that foreground motivation, embodiment, or the affective and values-based dimensions of learning — could yield different insights. The six analytical dimensions are my own construction. While consistent with established categories in learning theory, they would benefit from independent validation. Additional dimensions — embodiment, affect, the role of values and beliefs in shaping how learners engage with knowledge — could extend the analysis in important directions.

The HPE framing may limit perceived transferability. The theoretical foundations are not healthcare-specific, and the principles likely apply to other forms of professional education where practitioners navigate complexity and uncertainty. But this paper does not make that broader claim.

Conclusion

Health professions education faces a challenge that AI has made visible but did not create: we have prioritised content transmission and artefact production over the conditions under which professional learning actually occurs. The six principles presented here — dialogic knowledge construction, critical consciousness, adaptive expertise, contextual authenticity, metacognitive development, and networked knowledge building — offer a theoretically grounded response. They are not a defensive reaction to AI, nor an enthusiastic adoption of it. They describe what effective learning environments look like according to multiple established theoretical perspectives, and they provide guidance for ensuring AI integration supports these conditions rather than undermining them.

This framework is one map through a complex problem space, shaped by particular theoretical commitments and a particular analyst’s professional experience. But it is grounded in how learning works rather than in what AI can currently do, and that grounding gives it a stability that capability-dependent frameworks cannot offer. Empirical validation is the necessary next step — through design-based research that uses the principles to structure AI-integrated learning activities, comparative studies examining whether principled integration produces different outcomes than ad hoc adoption, and longitudinal work tracking whether the conditions these principles describe translate into the adaptive expertise that complex healthcare practice demands. The pedagogical question remains constant even as AI evolves: are we creating the conditions under which students develop the reasoning, judgement, and adaptive capacity that healthcare demands? These principles are a step toward answering that question — and toward building educational environments worthy of it.

References

Barnett, R. (2000). Realising the university in an age of supercomplexity. Society for Research into Higher Education & Open University Press.

Biesta, G. (2015). What is education for? On good education, teacher judgement, and educational professionalism. European Journal of Education, 50(1), 75–87.

Biggs, J. (1993). From theory to practice: A cognitive systems approach. Higher Education Research & Development, 12(1), 73–85.

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the real world (pp. 56–64). Worth Publishers.

Bleakley, A. (2010). Blunting Occam’s razor: Aligning medical education with studies of complexity. Journal of Evaluation in Clinical Practice, 16(4), 849–855.

Chee, J., Ahn, S., & Lee, J. (2025). A competency framework for AI literacy: Variations by different learner groups and an implied learning pathway. British Journal of Educational Technology, 56(5), 2146–2182. https://doi.org/10.1111/bjet.13556

Chi, M. T. H. (2009). Active-constructive-interactive: A conceptual framework for differentiating learning activities. Topics in Cognitive Science, 1(1), 73–105.

Dall’Alba, G., & Barnacle, R. (2007). An ontological turn for higher education. Studies in Higher Education, 32(6), 679–691.

Dawson, P., Bearman, M., Boud, D., Hall, M., Molloy, E., Bennett, S., & Joughin, G. (2024). Assessment might need to change just a little, or a lot: A psychometric perspective on assessment and generative AI. Assessment & Evaluation in Higher Education, 49(8), 1127–1139.

Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268.

Engeström, Y. (2001). Expansive learning at work: Toward an activity theoretical reconceptualization. Journal of Education and Work, 14(1), 133–156.

Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2024). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544

Fraser, S. W., & Greenhalgh, T. (2001). Coping with complexity: Educating for capability. BMJ, 323(7316), 799–803.

Freire, P. (2000). Pedagogy of the oppressed (30th anniversary ed.). Continuum.

Frenk, J., Chen, L., Bhutta, Z. A., Cohen, J., Crisp, N., Evans, T., Fineberg, H., Garcia, P., Ke, Y., Kelley, P., Kistnasamy, B., Meleis, A., Naylor, D., Pablos-Mendez, A., Reddy, S., Scrimshaw, S., Sepulveda, J., Serwadda, D., & Zurayk, H. (2010). Health professionals for a new century: Transforming education to strengthen health systems in an interdependent world. The Lancet, 376(9756), 1923–1958.

Glaser, B. G., & Strauss, A. L. (1967). The discovery of grounded theory: Strategies for qualitative research. Aldine.

Goldie, J. G. S. (2016). Connectivism: A knowledge learning theory for the digital age? Medical Teacher, 38(10), 1064–1069.

Head, B. W., & Alford, J. (2015). Wicked problems: Implications for public policy and management. Administration & Society, 47(6), 711–739.

hooks, bell. (1994). Teaching to transgress: Education as the practice of freedom. Routledge.

Illeris, K. (2003). Towards a contemporary and comprehensive theory of learning. International Journal of Lifelong Education, 22(4), 396–406.

Irby, D. M., Cooke, M., & O’Brien, B. C. (2010). Calls for reform of medical education by the Carnegie Foundation for the Advancement of Teaching: 1910 and 2010. Academic Medicine, 85(2), 220–227.

Jabareen, Y. (2009). Building a conceptual framework: Philosophy, definitions, and procedure. International Journal of Qualitative Methods, 8(4), 49–62.

Jonassen, D. H. (1995). Computers as cognitive tools: Learning with technology, not from technology. Journal of Computing in Higher Education, 6(2), 40–73.

Mishra, P., & Koehler, M. J. (2006). Technological pedagogical content knowledge: A framework for teacher knowledge. Teachers College Record, 108(6), 1017–1054.

Mitra, S. (2003). Minimally invasive education: A progress report on the “hole-in-the-wall” experiments. British Journal of Educational Technology, 34(3), 367–371.

Morrow, W. (2009). Bounds of democracy: Epistemological access in higher education. HSRC Press.

Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, 100041.

Plsek, P. E., & Greenhalgh, T. (2001). The challenge of complexity in health care. BMJ, 323(7313), 625–628.

Rittel, H. W. J., & Webber, M. M. (1973). Dilemmas in a general theory of planning. Policy Sciences, 4(2), 155–169.

Sfard, A. (1998). On two metaphors for learning and the dangers of choosing just one. Educational Researcher, 27(2), 4–13.

Siemens, G. (2005). Connectivism: A learning theory for the digital age. Journal of Instructional Technology and Distance Learning, 2(1), 3–10.

Stahl, G., Koschmann, T., & Suthers, D. (2006). Computer supported learning: An historical perspective. In The Cambridge handbook of the learning sciences (pp. 409–426). Cambridge University Press.

Suthers, D., & Rosen, D. (2011). A unified framework for multi-level analysis of distributed learning. In Proceedings of the 1st International Conference on Learning Analytics and Knowledge (pp. 64–74). ACM.

UNESCO. (2023). AI competency frameworks for teachers and students.

Van Der Vleuten, C. P. M. (2016). Revisiting ‘Assessing professional competence: From methods to programmes’. Medical Education, 50(9), 885–888.

Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

Warschauer, M. (2004). Technology and social inclusion: Rethinking the digital divide. MIT Press.

Wegerif, R., & Casebourne, I. (2025). A dialogic theoretical foundation for integrating generative AI into pedagogical design. British Journal of Educational Technology. Advance online publication. https://doi.org/10.1111/bjet.70026

Wong, G., Greenhalgh, T., Westhorp, G., Buckingham, J., & Pawson, R. (2013). RAMESES publication standards: Meta-narrative reviews. BMC Medicine, 11(1), 20.

Xu, W., Zhao, W., Li, Y., Qiao, L., Tao, J., & Liu, F. (2025). Enhancing self-regulated learning and learning experience in generative AI environments: The critical role of metacognitive support. British Journal of Educational Technology, 56, 1842–1863. https://doi.org/10.1111/bjet.13599