Rebuilding Surgical Training Around AI
Surgical training still runs on a century-old model.
A resident's growth is judged by whichever faculty member happened to be in the room that day.
A new Perspective in npj Digital Surgery argues AI can finally replace that guesswork with something continuous.
Three AI Domains, Long Treated Separately
The authors, from Harvard Medical School and Mass General Brigham, group current tools into three domains.
Large language models can run unlimited oral-board-style drilling, available at any hour without needing a senior surgeon in the room.
Computer vision and augmented reality can turn every operation into a teaching event, flagging safe and dangerous dissection zones in real time and generating feedback after the case ends.
Automated skills assessment can track instrument motion and economy of movement, scoring technical proficiency at a level approaching expert human raters.
Each domain already exists in some working form; the paper's contribution is arguing they should stop operating in isolation.
The Real Proposal: One Longitudinal Record
The authors' central claim is that no single tool matters as much as combining all three.
Cognitive reasoning from the language models, intraoperative judgment from the video systems, and technical execution from the skills scoring would feed one continuous record per trainee.
That record could flag specific knowledge gaps as they emerge, rather than surfacing them incidentally at a periodic review.
It would also let smaller, community-based programs offer a depth of feedback that has historically only existed at the largest academic centers.
The Deskilling Question the Authors Don't Dodge
The paper does not treat this as a purely upside story.
It cites a multicenter study showing endoscopists lost skill after growing reliant on AI during colonoscopy, and names that risk directly.
Training built around AI augmentation, the authors argue, still has to produce surgeons who can operate and reason without it.
Their proposed fix is designing systems that teach the process of surgical reasoning, not systems that simply hand trainees the answer.
What Has to Happen Before Any of This Is Real
The authors are careful to frame this as prospective, not deployed.
Automated assessment tools are only as good as the human evaluations they were trained against, and human raters carry their own inconsistency and bias.
Governance of intraoperative video data remains, in their words, an open problem.
Until validated against real educational and operative outcomes, they argue these systems belong in a decision-support role, not as a standalone judge of readiness.
