AI in Medical Education: Responsible Integration Across the Training Continuum

AI in medical education is moving faster than the rules for using it well. AI is now entering every stage of training, from admissions screening to continuing education for practicing physicians, yet guidance for responsible integration remains thin. A new Perspective in npj Digital Medicine, published in partnership with Seoul National University Bundang Hospital and currently in press, maps how AI should be matched to educational purpose across undergraduate (UGME), graduate (GME) and continuing medical education (CME).

The nine-author team, which includes Rushi Patel, Ceyhun Haziroglu and corresponding author Azra Bihorac of the AI for Health Institute at the University of Florida, argues that integration should reflect task structure, verifiability, consequence and oversight. They chose a conceptual synthesis over a quantitative review because the evidence is heterogeneous, changing quickly and unevenly spread across training stages.

Three domains for responsible integration

The Perspective treats AI as a supervised support layer that works across three connected domains. Institutional systems cover admissions, educational administration, learner monitoring and workforce-linked logistics. Learning systems include tutoring, simulation, feedback, assessment and clinical reasoning support. Readiness and safety systems cover AI literacy, competency development, ethics, governance, privacy, bias mitigation and professional accountability.

Ad

The domains overlap, but each has its own objectives, stakeholders and evidence requirements. Institutional systems support organizational decisions. Learning systems directly shape teaching, learning and assessment. Readiness and safety systems prepare learners and clinicians to use AI safely in practice.

What the early evidence shows

Admissions is where the effects on fairness and authenticity are most visible. AI screening of undergraduate applications has been evaluated for structured review and may improve efficiency and consistency. A small study found no significant difference in ratings between AI-assisted, AI-only and human-written application essays (Choi et al., 2025). Faculty struggle to tell authorship apart as well. In a multicenter study of orthopedic residency letters, faculty correctly identified 40% of the human-authored letters and 63% of the AI-generated ones (Simister et al., 2025). Detection tools are not a clean fix either. A small evaluation of three commercial tools gave variable classifications of human-written personal statements, which points to a real risk of false-positive attribution (Cumbo et al., 2025).

Learning systems show a clear progression across the continuum. In UGME, tutoring, virtual patients and assessment tools are generally well received. A preprint suggests standalone systems may not consistently improve diagnostic accuracy (Dao et al., 2022), and a comparative study found poorer results despite faster completion (Montagna et al., 2025). In GME, AI-assisted diagnostic systems improved residents’ accuracy on structured image-classification tasks. The authors caution that this reflects immediate supported performance, not durable clinical competence (Chi et al., 2026). For a related look at how AI is used to build assessment content, see our coverage of AI-generated medical exams. A specialty-specific tool also improved diagnostic performance in one comparative evaluation, although the benefit varied with expertise and carried a risk of overreliance (Rudie et al., 2021).

CME is the least developed part of the evidence base. For CME, AI has mainly been described in decision-support and point-of-care information systems rather than in purpose-designed educational interventions. Evidence on retention, behavior change and sustained professional learning is limited (Gil et al., 2019; Simon et al., 2019). A systematic review of explainable AI in clinical decision support found that evaluations focused mainly on predictive accuracy and subjective measures such as usability, acceptance and trust. It identified few clinician studies of overreliance or of responses to erroneous recommendations (Antoniadi et al., 2021).

Clinicians can be swayed by bad advice

Clinicians, including experienced ones, can be swayed by incorrect human and AI advice, and their accuracy declines when the system errs (Gaube et al., 2021; Tschandl et al., 2020). The authors argue that medical education should move past generic oversight language toward deliberate exposure. Curricula could present AI outputs of varying reliability, including seeded errors, and then test whether learners can detect, override and report bad recommendations. The educational target is calibrated trust, not maximal trust.

Five proposed competency domains for CME

The authors propose five competency domains for AI-focused CME. They are priorities drawn from the literature rather than validated standards, and the evidence behind most of them is still thin.

Source: Patel, Haziroglu et al., npj Digital Medicine (in press), Table 1. The domains are proposed priorities, not validated competency standards.

Gaps in readiness and governance in AI in medical education

Readiness and safety systems remain fragmented across the continuum. For a parallel view on who can reach these tools, see our piece on access to AI in medical education. Curricula differ in scope, duration, delivery and assessment. Implementation barriers include limited faculty expertise, uneven access across specialties and institutions, and weak links to educational governance and accreditation. National standards for AI literacy are underdeveloped, and AI competencies rarely feature in high-stakes assessment.

The authors argue that a coherent system should define progressive competencies and safeguards for supervised use. It should also specify permitted data uses, access to learner analytics, review and error-reporting processes, and ways for learners to challenge algorithmic judgements. Admissions algorithms need fairness monitoring, because models trained on historical decisions may reproduce existing inequities. Parallel questions arise in health technology assessment, and our coverage of AI governance in HTA covers them in more detail.

What future evaluation should measure

Future work should move beyond proof-of-concept studies. Priority outcomes include skill retention after AI is withdrawn, learner responses to incorrect output, subgroup fairness, privacy safeguards, faculty workload, transfer to clinical practice, and unintended effects such as overreliance and deskilling. The authors favor modular tools with defined functions over opaque end-to-end systems, because smaller tools can be validated, monitored and revised more transparently.

Post-deployment surveillance should track discrimination, measured as AUROC, and calibration across demographic and clinical subgroups. It should also track calibration drift on a defined cadence, clinician override rates and reasons, and periodic fairness audits tied to institutional governance review.

Why CME is the leverage point

Human-AI collaboration sits at the center of that argument, and our piece on collaborative intelligence in healthcare explores the same idea. The authors single out CME because attending physicians shape how AI is modeled, supervised and adopted in clinical learning environments. CME adoption is also likely to shape trainees’ attitudes and behavior. Evaluation should therefore measure demonstrated safe use and behavior change, not participation or perceived knowledge. The Perspective concludes that AI should augment rather than replace clinical judgement, and it remains subject to human accountability. Read the full paper at doi.org/10.1038/s41746-026-03362-2.