With artificial intelligence becoming increasingly present in health care, a recent scoping review of 54 studies from 22 countries found that recommendations for AI training in undergraduate medical education remain fragmented and heterogeneous. 

The review, published in npj Digital Medicine and led by Victor M. Hunt, MSc, of The Warren Alpert Medical School at Brown University in Providence, Rhode Island, and colleagues, conducted a PRISMA-ScR–guided search of PubMed, Embase, Web of Science, and ERIC from database inception through July 28, 2025. Of 4,071 records identified, 54 studies met inclusion criteria.  

Researchers extracted competency-related text and synthesized 564 eligible statements into 7 domains, 37 competencies, and 170 learning objectives. 

The 7 domains were Theory and Foundations of AI; AI Ethics; AI Law and Regulation; AI Professionalism in Healthcare; Clinical Applications of AI; Critical Appraisal of AI Output; and Research and Innovation in AI. 

Theory and Foundations of AI was the most frequently represented domain (46 studies), followed by AI Ethics (37), Critical Appraisal of AI Output (35), and Clinical Applications of AI (34). Research and Innovation in AI was the least represented domain with only 13 studies. 

Within the Theory and Foundations of AI domain, the most frequently represented competencies were “AI Foundations and Core Concepts,” “Data Science and Health Data Fundamentals,” and “Machine Learning Concepts and Paradigms.” 

In AI Ethics, the most frequent competencies were “Responsibility, Transparency, and Patient Rights” and “Bias and Equity.” For Critical Appraisal of AI Output, the most frequent were “Opportunities and Limitations of AI Systems” and “Analysis, Data Quality and Methodological Fit.”  

The researchers reported that competency proposals were largely derived from editorial or opinion-based literature (44%) while 67% of studies described theoretical or proposed competencies rather than implemented curricula. Only 6 studies reported pilot testing of curricula, and 8 evaluated educational outcomes. Across all included sources, 50% recommended mandatory AI curricula, 24% recommended elective instruction, and 26% did not specify implementation status.

The researchers noted several limitations, including the absence of formal risk-of-bias assessment, substantial heterogeneity in how studies defined competencies, and reliance on literature that was largely theoretical rather than outcome-based. They also cautioned that prevalence of competency themes reflects publication emphasis rather than educational priority. 

“This work underscores the need for collaboration among key stakeholders to refine, prioritize, and validate AI-related competencies across the diverse educational contexts of the AI-era undergraduate medical student,” the researchers wrote. 

Researcher Jonathan H. Chen, MD, reported research funding from multiple academic and federal sources, is a co-founder of Reaction Explorer LLC, which develops and licenses organic chemistry education software, and reported receiving expert witness fees and honoraria. The remaining researchers reported no competing interests. 

(Editor's note: Before final publication, the study will undergo further editing, and there may be errors present which affect the content, and all legal disclaimers apply.) 

Expert Insight

AACE Endocrine AI invited researcher Weston de Lomba, an MD candidate at the Warren Alpert Medical School at Brown University in Providence, Rhode Island, to elaborate on the findings.

Why does this study matter?

Increasingly, we are seeing artificial intelligence (AI) move into everyday clinical care, both at the systems level through EMR integration and at the provider level through changes in individual practice. These changes are invariably proceeding more rapidly than medical education has adapted. This study matters because it helps to clarify what future physicians should actually know about AI by organizing the fragmented literature into a structured set of proposed competencies for medical student education.

What data stood out to you?

We were pleasantly surprised that, overall, the literature did not primarily emphasize coding or technical development skills. There remains a persistent perception amongst many authors that their own work is pushing back against some idea that medical students need to learn the technical details of AI as if that were the dominant expectation. But across the literature as a whole, this is already the prevailing view—most authors do not argue that students should become AI developers, but rather that they should understand AI well enough to critically evaluate, interpret, and use it responsibly in clinical care. We were also struck by how much of the literature remains proposal driven. There are few truly rigorously evaluated curricula that have been practically implemented.

How might the findings influence endocrinology clinical practice?

The findings support a pragmatic approach to AI education that prepares these future physicians to use AI tools safely as they enter practice. In that sense, the immediate goal is not to withhold adoption until every trainee has deep technical expertise, but rather to ensure that learners are up to speed on the minimum competencies needed for responsible use. These competencies, in particular, emphasize interpreting outputs, recognizing bias and limitations, and knowing when human judgment should override the tool. Over time, more technical competencies may become increasingly important for physicians, but that is unlikely to be the most realistic immediate expectation for all medical students. In endocrinology, where AI may increasingly influence risk prediction, imaging, documentation, and decision support, these pragmatic competencies may be especially important for safe and effective adoption now.

Is there anything else you'd like to say about the findings?

If there is a single take-home message, it is that the field of medical education has moved beyond open-ended enthusiasm for AI and now requires practical educational validation. AI belongs in medical education, but the next step is to refine, prioritize, and test these competencies with educators, students, clinicians, and other stakeholders so that implementation remains thoughtfully evidence-based across feasible curricula.