Familiar but unprepared: Artificial intelligence training needs in graduate medical education

Graduate medical education trainees increasingly use artificial intelligence tools without formal training. This study quantifies the gap between artificial intelligence exposure and readiness and examines attitudes, preferences, and barriers to guide curriculum development.

In 2025, a 4-week cross-sectional survey, consisting of 5-point Likert scale questions, of graduate medical education trainees at a single academic medical center assessed artificial intelligence familiarity, confidence, attitudes, curriculum preferences, and barriers to training. Descriptive statistics, χ² tests, Spearman correlations, Jaccard similarity analysis, and network analysis characterized response patterns.

Of 149 eligible trainees, 69 (48.3%) responded. Of all respondents, 85.5% (n = 59) reported active artificial intelligence tool use, and 2.9% (n = 2) had received formal artificial intelligence instruction. Median familiarity (3 [interquartile range, 3-4]/5) exceeded median confidence (3 [interquartile range, 2-3]/5) (P = .02), with 38.5% of high-familiarity trainees reporting low confidence. Trainees uniformly preferred postgraduate year-1 introduction (65.2%) and longitudinal curricula (62.3%), with no differences by training level or specialty. Barrier co-occurrence analysis revealed institutional support and faculty expertise deficits clustered strongly (Jaccard = 0.74). Network analysis identified efficiency beliefs as the central attitudinal hub.

Graduate medical education trainees demonstrated high artificial intelligence exposure but low confidence, with formal training being rare. The familiarity-confidence gap suggests that passive exposure does not confer confidence. Uniform curricular preferences support standardized, cross-specialty implementation. Addressing institutional support and faculty expertise together, while separately targeting time constraints, may optimize artificial intelligence curriculum adoption.