Surgery ·
Surgeons Accepted 61% of Incorrect AI Guidance in Robotic Surgery Experiment
A controlled experiment involving 19 surgeons tested responses to mock AI overlays on robotic cholecystectomy images. Participants followed incorrect guidance in 60.6% of frames where the overlay was deliberately unsafe, while confidence increased after exposure to AI.
- 19 surgeons tested mock guidance on robotic cholecystectomy images.
- Incorrect suggestions were accepted in 60.6% of relevant frames.
- Confidence increased after exposure to AI overlays.
- Static-image experiment did not use an actual AI model.
In a two-phase controlled experiment, 19 surgeons (10 residents, two fellows and seven attending surgeons) marked incision lines on 20 static frames from robotic cholecystectomy videos. They first worked without guidance, then viewed mock AI overlays, which were correct or deliberately shifted toward unsafe anatomy. No AI model was used. The primary outcome was the change in incision-line position; secondary measures included acceptance of guidance and confidence.
Across 231 paired observations, mean line displacement did not significantly differ between correct and incorrect overlay conditions (75.5 vs 71.1 pixels; P=.35). Surgeons accepted 43 of 71 incorrect suggestions (60.6%) and 123 of 160 correct suggestions (76.9%); overall behavioral accuracy was 65.4%. Self-reported confidence rose after AI exposure (change +0.65; P<.001).
The findings raise a safety concern about relying on surgeon oversight to catch erroneous intraoperative guidance. However, the experiment used static images and mock overlays rather than a functioning AI system or live surgery, and the sample was small. The results do not establish how surgeons would respond in clinical practice, but support evaluating human-AI interactions before deployment.
In practice
In this small controlled experiment, surgeons accepted incorrect mock AI guidance in 60.6% of frames containing deliberately unsafe suggestions, and reported higher confidence after AI exposure. The findings do not support changing clinical practice: participants assessed static images, no functioning AI system or live operation was involved, and the sample included only 19 surgeons. They indicate that human oversight should not be assumed to prevent errors and that responses to intraoperative AI need evaluation in more realistic, prospective settings before clinical deployment.
AI-generated clinical interpretation · Not editor-reviewed
Is this summary clinically accurate?
Help fellow clinicians: your rating sends inaccurate summaries straight to our editors.
Sign in to rate this summary →Source
Surgical Endoscopy: The surgeon in the loop: automation bias and human–AI interaction in robotic surgery ↗This is an automated AI-condensed summary that has not yet been reviewed by an editor. Always consult the full item at the original source.