Surgical training in the era of artificial intelligence and large language models: a narrative review from the perspective of a surgical resident

Abstract

Artificial intelligence (AI) and large language models (LLMs) have entered surgical residency faster than the curricula, assessment systems and governance frameworks designed to accommodate them, yet most published appraisals adopt the standpoint of educators rather than that of the trainee, who is simultaneously the principal user and the principal subject of these tools. This narrative review, informed by the SANRA criteria, searched five databases from January 2015 to June 2026 and synthesises the evidence from the resident's perspective. Genuine educational value is supported in four domains: on-demand knowledge synthesis and examination preparation, intelligent tutoring and automated feedback in simulation, objective video- and kinematics-based skill assessment, and relief from administrative burden. Randomised trials show that AI tutoring improves machine-derived performance metrics but that gains are often not reproduced on blinded human ratings, and that AI combined with expert instruction outperforms AI alone. Against this stand substantial risks: hallucination and guideline drift, with accuracy falling sharply on current, guideline-specific content; the first real-world signal of clinical deskilling following routine AI exposure in a procedural specialty; cognitive offloading, with attendant mis-skilling and never-skilling during formative training; threats to assessment construct validity and to the apprenticeship relationship; breaches of confidentiality; algorithmic bias; and a widening divide between well-resourced and resource-constrained programmes. Surveys show substantial trainee uptake alongside negligible formal instruction. The governing principle for the trainee is sequence: independent reasoning and unassisted practice must precede augmentation. Training programmes should teach structured AI literacy, protect AI-free deliberate practice, and adopt verification-centred supervision.