Exploring the integration of artificial intelligence into residency education curriculum: qualitative interviews

Synthesis: Bursic and colleagues interviewed twelve physicians who lead AI work in postgraduate medical education across Canada and the United States, and report five barriers to putting AI literacy into residency curricula: curricula that cannot keep pace with the tools, thin faculty expertise, institutional and regulatory AI Governance that has not caught up, no national competency standards, and cultural resistance tied to professional identity. Their recommendation is not more content but durable competencies — frameworks for appraising AI output, bias awareness, and Ethics — around which programs attach modular, locally available tool work. Faculty development and co-learning with residents carry the load, because participants expected trainees to become more fluent than the educators. No curriculum or learner outcome was measured: this is a qualitative account of expert perception from a small, self-selected sample that excluded residents.

Key Findings

Participants described AI's pace as faster than any previous medical curriculum cycle, and argued that teaching durable frameworks for evaluating a tool beats teaching content about any single tool.
Faculty expertise was described as thin beyond a few known experts, with informatics units not yet mature enough to lead curriculum development; upskilling, co-teaching, and co-learning were the proposed remedies.
Most participants reported no institutional AI policy, or an incomplete one, and said data ownership, consent, and privacy remained unresolved; US institutions were described as ahead of Canadian ones.
No nationally endorsed AI competency standards existed. The Royal College of Physicians and Surgeons of Canada was seen as well positioned to convene and standardize, yet currently behind.
Access to AI tools, including widely adopted ones such as AI scribes, varied by health system and hospital, so what a resident could learn depended on their site's approvals and budget.
Professional identity and responsibility were described as unsettled by AI, and participants proposed ethics, bias mitigation, and patient safety as the ground on which trust is built.
The proposed structure is a modular curriculum: a core of AI and machine-learning literacy, critical appraisal, bias detection, metacognition for human-AI interaction, and ethics, with specialty and local-tool content reviewed every six to twelve months.

Curricular agility and durable competencies

Participants treated the mismatch between tool release cycles and curriculum review cycles as the primary design problem, not an inconvenience. Static content about a named product was described as obsolete before approval, so several argued for competencies that survive tool turnover: appraising an output, recognising when a tool does not fit the clinical task, and reasoning without assistance when necessary. That reframing moves the difficulty from keeping content current to specifying what remains stable, which is a curriculum design question rather than a technology one. It also creates an assessment problem the study does not resolve, since durable competencies are harder to certify than a completed module. Where a program already has informatics capacity, participants thought content could be added quickly; the lag they described is a feature of institutions without that capacity.

Faculty, co-learning, and the expertise gap

The sharpest asymmetry in the interviews ran between trainee fluency and educator fluency. Participants described residents arriving with more practical AI experience than many faculty, and framed that inversion as an opportunity rather than a threat: co-learning, co-teaching with informatics or engineering colleagues, and embedding ethicists in sessions on bias and accountability. Protected time and communities of practice were named as the conditions under which staff physicians could build the competence that AI competency frameworks assume they already hold. The study's own design leaves the trainee half of that pairing unexamined, so the co-learning recommendation rests on faculty perception of learner readiness rather than on learner report.

Governance, standards, and infrastructure

Regulatory and medico-legal uncertainty recurred in every interview, usually through privacy: patient information entering a tool whose handling nobody had approved. Participants described policies as arriving slowly and unevenly, and infrastructure as determining exposure — a resident at a site without an approved AI scribe cannot be taught to work with one. The standards gap compounds this, because without national expectations each program improvises, and the participants who wanted guidance also predicted that any published standard would be a year behind the field. The tension they described is real: standards published early are stale, standards published late leave practice ungoverned, and the alternative they favored was a national body that curates expertise continuously rather than issuing a document.

Identity, trust, and what to teach

The cultural theme is the least tractable. Participants described colleagues who saw AI as an extension of practice and colleagues who read it as a challenge to what a physician is, captured in one remark that the professional identity medical education spends years forming was "being changed overnight". Responsibility surfaced repeatedly: a clinician using a tool may have less control than before, yet the oath and the accountability do not move. Teaching grounded in ethics, bias mitigation, and patient safety was offered as the way to make adoption defensible, and participants noted that institutional ethics boards were themselves struggling to evaluate AI projects.

What this means for practice

Teach frameworks for appraising AI output rather than product-specific skills, and schedule a review of any tool-specific content at least annually.
Build a core module on AI literacy, critical appraisal, bias, and human-AI metacognition, then attach specialty and local-tool elements to it.
Fund faculty development and co-teaching explicitly; treat resident fluency as a teaching resource rather than a gap.
Ask for institutional policy on data handling and patient privacy before designing clinical AI exercises, and record what a site's approvals permit.
Involve ethicists and medico-legal experts in curriculum design, not only in incident review.

Limitations

Twelve participants, drawn from faculty and educational leaders already interested in AI, mostly internal medicine; specialties where AI is more mature, such as radiology and pathology, are underrepresented.
Residents and fellows were not interviewed, so trainee readiness, enthusiasm, and concern are reported second-hand.
The sample is Canadian and selected US contexts, which limits transferability to other regulatory and health-system settings.
Perceptions, not outcomes: no curriculum was implemented and no learning was measured.

Connected Concepts

Connected Articles

Institutional approaches to artificial intelligence policy and guidance in health informatics and information management education: emerging trends and inconsistencies — Institutional AI policy in health informatics
From fear to innovation: A case study of transformative faculty development for ethical AI integration in higher education — Faculty development for ethical AI use
Designing faculty standards for technology integration in higher education institutions: a design-based research study — Faculty technology integration standards
When the algorithm enters the classroom: A critical integrative review of large language models, nursing education structural gaps, and the reconstitution of professional identity — Professional identity and LLMs in nursing education
AI as Teammate: Rethinking Task Distribution in Medical Training — Task distribution with an AI teammate in medical training
A Stratified Multi-Rater Evaluation of LLM-Based Virtual Standardized Patients with a Deployed Data-Generation Platform — Virtual standardized patients in OSCE evaluation
Mapping the Emerging Curriculum for AI-Assisted Software Engineering via Syllabus Analysis — AI in software engineering syllabi

Citation

Bursic, B., Sheth, U., Chopra, S., Sibbald, M., & Martin, L. (2026). Exploring the integration of artificial intelligence into residency education curriculum: qualitative interviews. BMC Medical Education. https://doi.org/10.1186/s12909-026-10684-y

Metadata

Study design	Interviews, Thematic Analysis
Discipline	Medical Education
Level	Higher Ed
Intended audience	Instructors, Administrators, Researchers
Learning and instruction	Workplace Learning
Research methods and evaluation	Qualitative Research
Institutions and policy	AI Governance, Educational AI Policy, AI Regulation in Education
Equity, ethics, and responsible use	Ethics, Bias Mitigation, Privacy, Trust
Foundations of AI in education	AI Literacy, Curriculum Design, Teacher AI Competency, Human AI Collaboration
Page kind	Synthesis

Source History Edit Chat