On this page

Medical and Health Professions Education (HPE) — the teaching and training of medical, nursing, pharmacy, and allied health professionals. AI is reshaping this domain through clinical simulation, reinforcement learning trainers, adaptive learning, and the application of foundational learning principles (experiential, situated, and distributed cognition) in health-professions contexts. Because HPE is high-stakes, competency-based, and clinically embedded, it raises distinct questions about AI's role in skill acquisition, patient safety, and the educator's judgment. Nursing, the most heavily studied program in HPE, is covered in full under nursing education, which develops the competence boundary, identity-formation, and workforce evidence that appears here only in outline.

Questions to Consider

  • In medicine, AI benefits like scalable practice and adaptive feedback must be balanced against risks like erosion of hands-on clinical skill — where errors carry direct patient consequences. Where would you draw the line between what AI should do and what a trainee must practice for themselves?
  • AI can now operationalize age-old learning principles such as experiential, situated and distributed cognition, and simulate or personalize practice. What makes the educator's judgment still indispensable?
  • Reinforcement-learning trainers and agentic AI are now used for clinical and procedural skills in residency. If an AI agent trains a resident on a procedure, how would you verify they've actually learned it safely before they do it on a patient?
  • Because health-professions education is high-stakes and competency-based, assessment questions carry particular weight. How might AI-assisted assessment both improve and threaten the evaluation of clinical competence?
  • Over-reliance on AI is a specific concern in medicine. How do you think training with AI could produce a clinician who is more confident but less able to reason independently — and what could guard against that?

Introduction

AI in medical and health-professions education is a growing strand of AI-in-education research. Unlike general higher education, HPE is oriented toward the development of clinical competencies, procedural skills, and professional judgment, which shapes how AI tools are designed and evaluated.

How AI appears in health-professions education

  • Operationalizing foundational learning principles. Fowlin et al. (developed at the Medical University of South Carolina) argue that AI should be used to operationalize age-old learning principles — Dewey's experiential learning, situated cognition, and distributed cognition — rather than replace the educator's guiding role. AI enhances personalized and adaptive learning while the teacher remains central to student engagement and outcomes.

  • Clinical skills and reinforcement learning. ResidencyRL uses reinforcement learning to train clinical reasoning and procedural skills in residency, demonstrating AI as a skills-training partner in real clinical workflows.

  • Simulation and agentic AI. Agentic AI simulation in brachytherapy shows how AI agents support hands-on procedural training in medical specialties.

  • Gamification of medical learning. MedGame applies gamification and LLMs to engage medical students.

  • Nursing and interdisciplinary education. AI transformation of nursing education documents how AI reshapes nursing curricula and instruction. Nursing is kept as one program among medicine, pharmacy, dentistry, and allied health, organized around the shared problem of scalable clinical competence; the nursing-specific record — a competence boundary where AI's documented benefits stop, an explicitly professional-identity framing, and students' workforce anxiety — belongs to nursing education and is cross-referenced here rather than restated.

  • AI-powered simulation in nursing. Jiang et al. (2026)'s mixed-methods systematic review of 19 studies (N = 1,253) finds AI-driven simulations (GenAI/LLMs, virtual patients/mannequins, AI-enhanced VR/MR, and chatbots) significantly improve cognitive knowledge and affective outcomes (self-efficacy, communication confidence) in the strongest designs, but are inconsistent for complex psychomotor skills — one RCT found AI-assisted simulation inferior to standardized patients. Their concept of an "authenticity gap" — a learner-perceived shortfall in emotional resonance, nonverbal cues, and tactile examination — grounds why AI is best for highly structured objectives (foundational communication, history-taking) and should sit in a stepped simulation continuum alongside, not instead of, human-standardized patients and clinical placement. This is a distinctive, evidence-anchored refinement of the domain's simulation strand and parallels the acceptance-trust finding that learners favor human feedback perceived as "benevolent" over AI seen as merely "competent."

  • LLMs and the reconstitution of professional identity in nursing (2026). Sun et al. (2026)'s critical integrative review of 489 studies across 47 countries (Whittemore & Knafl framework) shifts the question from what LLMs can do in nursing education to what their integration does to the developmental processes through which a nurse is formed. Framed through Benner's skill acquisition, cognitive load theory, automation bias, and Wenger's identity formation, they find the same technology both enhances and erodes competence depending on whether the displaced cognitive work is extraneous to, or constitutive of, the target competence — unstructured reliance produced measurable deficits in ethical reasoning and clinical judgment. Their Professional Identity Tension Model distinguishes LLM integration that supports professional formation from integration that silently substitutes for it, and Structural Empathy Suppression reframes apparent AI "outperformance" on relational metrics as a symptom of systemic overwork rather than genuine AI empathy. An evidence gap map shows the highest-policy-consequence domains (professional identity, relational/ethical competency, long-term outcomes) rest on the thinnest rigorous evidence — a caution against confident deployment claims in nursing education.

  • Multi-agent AI standardized patients in clinical interview training. In a 2026 randomized controlled trial with 95 analyzed medical students (Yang et al.), multi-agent AI standardized patient training raised final OSCE-aligned examination scores over a structured non-LLM progressive-disclosure control (71.8% vs. 55.6%; β = 16.4 percentage points; P = 3.30e-4) while leaving binary diagnostic accuracy statistically identical (84% vs. 86%; P = 1.000). The dissociation indicates that early clinical-training gains can surface first in consultation process quality — communication (mean 3.53 vs. 2.64 on a 1–5 OSCE scale; P = 4.50e-4), empathic expression (+31 percentage points on the "expressing empathy" checklist item), and selected history-taking behaviors — before any improvement in diagnostic endpoints, and that AI patients still need human raters and faculty for professionalism and readiness judgments.

  • AI decision support helps trainees most where they are weakest. Across 15 residents reading 186 prostate biopsy slides, AI assistance raised accuracy from 62.3% to 75.1% (+12.9 percentage points) for less experienced readers versus 77.9% to 83.0% (+5.1 pp) for experienced ones (Zhu et al. (2026)). The AI display converted 39% of 786 unaided errors to correct diagnoses but 3.9% of 1,913 correct ones to errors, and the fixed reading order leaves educational value a hypothesis for counterbalanced studies.

  • Criterion-level scoring localizes what an AI patient cannot do. Across three Korean OSCE cases, three of four backbones sat within 0.5 Likert points of an expert reference on comprehensibility, consistency and hallucination-freeness, yet weaknesses concentrated in clinical reasoning and patient education and in a Korean negative-form penalty on fluency (Yi, Park & Kim (2026)).

  • Integrating AI into a residency curriculum is a governance problem before it is a content problem (2026). Bursic et al. interviewed twelve physicians who lead AI work in Canadian and US postgraduate medical education. The five barriers they name are structural rather than pedagogical: review cycles slower than tool release, faculty expertise thin beyond a few experts, governance that has not caught up on privacy and data ownership, no national competency standards, and cultural resistance tied to professional identity (Bursic et al. (2026)). Their proposal is a modular core — AI literacy, critical appraisal, bias detection, human-AI metacognition, ethics — with specialty and local-tool content reviewed every six to twelve months. The clinical evidence documents what AI can teach; the curriculum question is whether a program can teach it at all.

  • Students default to treating an AI patient as a searchable database. Voice-interviewing a genAI avatar in a PBL pilot, students asked closed-ended, efficiency-focused questions — one asking whether the AI was "just, like, a question base" — and the encounters ran 55–65 minutes versus 36–39 for the legacy module (Mool et al., 2026).

  • Initiative is a second axis beside task allocation. Jin et al. (2026) randomized 79 medical and nursing students and graduates to a reactive agent that waits for prompts or a proactive one that drives sequenced questions, writing analytical pieces over visuals built from the multimodal traces of a four-person nursing simulation (a bar chart of time per behavior, a communication network, a ward heatmap). The two conditions differed sharply in epistemic structure (Cliff's delta of -0.56 and -0.78) while final scores did not, so where a tool sits in an allocation and how much initiative it demands are separate design decisions. A second governance gap sits on the production side. Policies in health-professions education overwhelmingly govern what learners do with AI and rarely what faculty produce with it, which Chew (2026) treats as structural: AI-generated images, cases and test items enter curricula as teaching content, where a plausible error reaches every learner and may never generate the corrective signal a student's mistake does. He sorts the failures into representational, reasoning and assessment types, and argues that consequence asymmetry — learner errors are correctable, curriculum errors propagate as established knowledge — justifies four institutional capabilities: knowing what is in use, testing before learners see it, acting on findings including withdrawal, and monitoring approved tools for drift. Audit capacity, on this account, is a potential new equity divide (Chew, 2026).

  • Cross-program analytics without shared student records. Kovor, Osei and Agyemang (2026) partitioned the Open University Learning Analytics Dataset into five program-discipline clients as health-professions analogues and found federated logistic regression kept 99.3-99.9% of centralized performance, so the obstacle to sharing a risk model across programs is organizational rather than statistical. The reason a pooled model fails is pedagogical heterogeneity: their illustration has a nursing program at 30-35% at-risk against 15-18% for a concurrent medicine program, and forum activity ranked last as a predictor in every discipline because peer learning in these programs happens in clinical placements.

  • Task allocation and the SCAN framework. Tsim et al. reframe AI integration from learner "misuse" to misclassification — a failure of real-time metacognitive evaluation. Their SCAN framework (Substitute, Complement, Aid, Non-Negotiable), grounded in Vygotsky's zone of proximal development, allocates generative AI tasks by the individual learner's developmental state and identifies passive engagement within AI-scaffolded tasks as a hidden pathway to mis-skilling that requires re-identification from AI to expert assistance with human epistemic auditors.

  • Human-in-the-loop instructional asset generation. Gen-Mentor (Dong et al. 2026) integrates a vision-language-model backbone into a dental-radiography workflow: Faster R-CNN localizes four target findings (Filling, Implant, Impacted Tooth, Cavity), a conditional diffusion model generates class-specific synthetic ROI candidates, a VLM produces evidence-linked captions, and an LLM reformats them into case descriptions, comparisons, and quiz prompts — all before structured expert review. Evaluated with 45 dental students (mean SUS 72.7), it demonstrates how human-in-the-loop review of AI-generated instructional assets can expand case diversity and immediate-feedback support while retaining expert oversight over what students see.

  • AI grading of pharmacy exams. Falahat, Das, Bhaumik & Thambi (2026) evaluated ChatGPT-5 against human faculty grading of a 21-item pharmacy exam (16 students) across multiple-choice, select-all-that-apply, fill-in-the-blank, listing, short-answer, and essay items. The model matched faculty closely on objective items (CCC 0.935–1.000) but was unreliable for listing, short-answer (CCC ≈0), and essay (0.341–0.854) responses, and a structured rubric did not consistently improve agreement — evidence that in high-stakes, competency-based assessment in health professions, AI suits well-specified items while human review remains necessary for subjective, open-ended clinical reasoning.

  • A validated LLM judge for clinical communication. Hasan et al. (2026) scored the 3E communication skills from transcripts alone with a judge that reached Pearson 0.759 and ICC(A,1) 0.746 against the consensus of human raters, inside the spread of individual humans, though its weakest dimension (Be Explicit) did not improve across encounters. Both lines of work measure communication behavior rather than self-report; empathy collects what is known about teaching and assessing the construct they serve.

  • GenAI in scenario-based healthcare education. Neto and colleagues (2026) systematically reviewed 23 studies of GenAI across scenario-, case-, problem-, and simulation-based learning in healthcare education (PRISMA 2020). Their central finding is that prompt design functions as instructional specification — encoding the cognitive targets and quality criteria implicit in expert authoring — yet only 34.8% of studies aligned generated content with instructional frameworks and only 34.8% reported prompting in enough detail to reproduce. GPT-4 dominated implementations (44.4%), hybrid human-AI collaboration outperformed fully automated approaches, and evidence was strongest for higher-order cognitive skills but inconsistent elsewhere. This grounds scenario/simulation-based and problem-based pedagogy in medical education with validation- and integration-quality benchmarks.

  • AI scoring of open-ended exam questions. Olvet et al. (2026) tested whether GPT-4 could reliably score open-ended questions on pre-clerkship assessments at two US medical schools. With faculty iteratively refining scoring rubrics across three rounds of error-pattern analysis, AI–faculty inter-rater reliability reached substantial-to-almost-perfect agreement on three of four questions (weighted kappa up to 0.94) but only moderate on the holistic-rubric item (κw = 0.54); discrepancies traced to both raters (GPT-4 over-scoring multiple-answer or rubric-absent-vocabulary responses; faculty being overly generous) and occasional feedback inaccuracies keep humans in the loop. The authors argue the case for automated OEQ scoring is strengthened because ~82% of US medical schools grade pre-clerkship work pass/fail, where exact AI score agreement is not always required.

  • Repeated runs do not reproduce the same grade. In a master's-level public health course, four general-purpose LLMs graded the same 32 anonymized essay submissions as two human examiners; ChatGPT in thinking mode reached weighted kappa 0.718, but regrading identical submissions on five different days reached only Krippendorff's alpha 0.625 (Brevik et al. (2026)). The authors advise averaging multiple assessments for high-stakes use.

  • Evaluating AI teaching agents, not just deploying them. Zhang et al. (2026) deployed eight LLM teaching agents across four role-play paradigms (patient, student, expert, family member) in an endocrinology curriculum and scored 167 student dialogues with an 8-dimension teaching-quality rubric. Platform scores diverged sharply from rubric quality; agents differed most on knowledge dimensions and least on role enactment; adaptive difficulty calibration was a shared weakness; and strengthening an agent's empathy raised role-play quality without improving knowledge coverage. The finding that role-play paradigms can be separated (beyond the default patient-doctor script) points to designing and evaluating agents against explicit pedagogical dimensions.

  • The shape of the field's own literature. Sriram, Nichols, Ganti and Gue (2026) map the ethical-AI-in-medical-education literature itself: 1,403 Web of Science publications from 1995 to 2026, negligible for two decades, then 113 in 2023, 256 in 2024 and 500 in 2025; 39 countries clear the authorship threshold, led by the United States (533 publications) over China (189) and England (95); and the most-cited works are capability tests — ChatGPT's performance on medical licensing examinations above all — rather than governance scholarship. Their reading is that the field has grown fast without growing evenly, and that empirical evaluation of governance frameworks is the work the citation record is not rewarding. It is a useful caution for anyone treating the volume of published AI-in-medical-education research as evidence that its safe-use questions have been answered.

A 153-report scoping review of GenAI in medical education maps applications — simulation and clinical skills largest (47 reports), ahead of assessment generation and feedback (22) and case-based learning and clinical reasoning (21) — yet only 13 reports carried a follow-up or retention signal and no patient-level outcome was found (Zhao et al. (2026)).

  • AI-powered PBL and CBL: six randomized trials, two of them CBL. Wei et al. (2025) pooled four trials for knowledge (46%, 95% CI 0.18 to 0.73), three for satisfaction (0.7, 95% CI 0.47 to 0.92), and could not pool clinical reasoning because the instruments differed. Only two trials tested case-based learning and the review did not analyze the formats separately, so the result supports AI-supported problem designs without identifying which case features matter.

Why it matters

HPE is a high-stakes, competency-based domain where AI's benefits (scalable practice, adaptive feedback, simulation) must be balanced against risks (Over-Reliance, erosion of hands-on clinical skill, ethical and safety concerns). The general concepts — teacher role, assessment, feedback, equity, and ethics — apply with particular intensity in health professions, where the consequences of error reach patients directly.

Wang and Shan (2026) name the resulting "safety gap" — the divergence between a student's AI-assisted performance and their unassisted ability to verify it — and prescribe withholding solutions, constructive cognitive friction, and Socratic or Adversarial AI architectures for high-stakes clinical training.

  • What the curricula teach about data provenance. Giancotti et al. (2026) coded the publicly advertised curricula of 36 U.S. health AI programs and found ethics covered in 29 (80.6%) but data provenance at a structural or dataset-limitation level in only 8 (22.2%), with no program addressing the environmental cost of computation and none of the seven master's degrees meeting the provenance threshold. Of 92 named instructors, 76.1% came from computational disciplines and 3.3% from the social sciences or humanities. The gap is clinically consequential: the deployment failures the paper opens with were introduced upstream, in how training data were generated or labeled, which is a teachable competency the curriculum appears not to carry.
  • Use has outrun instruction. Among 310 health science students, 90% had used ChatGPT while 82.6% had received no formal AI instruction; usage frequency tracked higher GPA, and 61% reported ethical concerns (Alshahrani et al., 2026).

The sub-Saharan African evidence makes that gap concrete under material constraint. Surveying 417 undergraduates across six allied-health programs at one Ghanaian public university, Ago et al. (2026) found 76.3% (n = 318) had received no prior teaching in responsible AI use, and students named paid subscriptions and internet connectivity among the constraints on their use. What they asked for was curriculum rather than prohibition: AI literacy and capacity building drew 185 codes, over three-quarters of suggested support strategies.

Implications for health-professions educators

  • Use AI to operationalize learning principles, not replace the educator. Fowlin et al. argue AI should operationalize experiential, situated, and distributed-cognition learning while the teacher remains central to engagement and outcomes.
  • Leverage AI for clinical skills training. ResidencyRL and agentic simulation show AI as a skills-training partner in real clinical workflows — embed it where it adds safe, scalable practice.
  • Balance high-stakes benefits against over-reliance. HPE is competency-based and high-stakes; guard against AI substituting for hands-on clinical skill and judgment, and apply feedback, assessment, and ethics considerations with particular care.
  • Adapt gamified and interdisciplinary AI thoughtfully. Gamified LLM learning and nursing-education transformation show promise but need evaluation for safety and skill outcomes; for nursing specifically, that safety-and-skill evidence — including the RCT in which AI-assisted simulation underperformed standardized patients — is gathered on nursing education.
  • Order the curriculum by the evidence, not by the tools on hand. The same KAS map shows where the evidence is thick and thin: ethics knowledge (14 studies), responsibility (14), application skills (16) and evaluation skills (18) carry consensus, while technical knowledge, application knowledge and operational skills sit at 9 each and collaboration at 7. Where infrastructure is unreliable, reach the technical layer through screen-based cases and open-access model demonstration rather than waiting for high-fidelity simulation.

Common Misconceptions

  • "Simulation benefits are established in medical education, so AI simulation will too." A mixed-methods systematic review of 19 studies (N = 1,253) finds AI-driven simulations produce effects in that literature — a modest base, and 1,253 participants across 19 studies is a narrow foundation for a technology being deployed at scale.
  • "The question is what LLMs can do in clinical training." A critical integrative review of 489 studies across 47 countries shifts the question from what LLMs can do to what they should be trusted with. The capability question is largely answered; the governance question is not.
  • "Standardized patients are too costly to scale." A randomized controlled trial with 95 analyzed medical students used multi-agent AI standardized patients in clinical interview training. The constraint being addressed is the availability of trained human actors, which is a staffing problem rather than a fidelity one.
  • "Adopting a published AI competency framework settles the curriculum question." Xiang et al. (2026) screened 1,750 records to 23 studies — 15 reviews, 4 qualitative and 4 quantitative, across 15 countries and 2022 to 2026 — and mapped them onto the Knowledge, Attitude and Skill model. Application domains concentrated in medical education (n = 11, 47.8%), general education (n = 9, 39.1%) and other education (n = 3, 13.0%), and the named nursing frameworks they compared — N.U.R.S.E.S., AI-ABCs and CARE — either lack clear pathways for pedagogical implementation or cover isolated skills without assessment methods. The 12-sub-dimension map they build is a proposal awaiting validation, which makes it something to teach against rather than a settled curriculum.

Graph-grounded teaching in laboratory courses

A 2026 quasi-experiment in a Chinese virology laboratory course rebuilt the module around a course knowledge graph, a graph-grounded retrieval assistant, adaptive routing with published thresholds, and virtual simulation, then compared 30 students with 30 in the unchanged course. The experimental class scored 11.4 points higher on the final theory examination (d = 1.39) and 13.3 points higher on the laboratory operation checklist (d = 1.61).(Liu et al., 2026)

Two cautions travel with the numbers: allocation was by intact class, so clustering is unmodelled and the intervals are likely too narrow, and the four components were introduced together, so the effect belongs to the package rather than to the knowledge graph or the assistant.

Connected Concepts

Connected Articles

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.