Surgical AI Maps Safe Zones and Drafts Notes as Safety Questions Grow
Researcher Madani leads research in surgical AI focused on safety and precision in the operating room. At the Toronto General Hospital Research Institute, the team has used AI to identify safe and dangerous zones to guide surgery. The work was profiled by the UHN Foundation on Dec. 8, 2025, and the zone-mapping approach was described by Hospital News on Feb. 28, 2024.
The main job is orientation during surgery, what doctors call intraoperative guidance. The system learns from surgical video to tell apart tissue areas, labeling where a surgeon can act and where risk is higher, like a navigation aid that colors the field. The guidance appears during the operation. The surgeon keeps control of tools and decisions.
That guidance task is different from a second use now in view. AI systems can watch video of an operation and then draft the post-surgery report, what doctors call the operative note, according to Reuters reporting on Feb. 14, 2025. In this case the input is video from the procedure. The output is structured text for the medical record.
On safety, a Reuters investigation published Feb. 9, 2026 reported botched surgeries and misidentified body parts as AI entered the operating room. The investigation did not find a single cause across cases. It listed system error, heavy reliance by users, and natural differences in anatomy as part of the same area of concern.
Reuters also reported on June 9, 2026 that AI is saving clinicians time, but a majority of health professionals say training in the technology is inadequate. Time saved on notes and workflow help is real. Confidence in safe use is lagging behind wider use.
The broader context here is a split path. One track points toward tighter control during surgery and less paperwork after. The other points toward errors that are hard to spot in real time. The open question is how these tools perform outside controlled tests, when views are poor, there is bleeding, scar tissue from past surgery that surgeons call adhesions, or unusual anatomy, a problem engineers call distribution shift.
Looking at what this means for deployment, the Toronto work shows where testing must focus. Zone labels need checking frame by frame and stage by stage, measuring false-safe errors and false-dangerous errors. Note drafts need checking for missing details, added details, and wrong side, what doctors call omission, commission and laterality. Both need audit trails that link what the model said to the video and to what the surgeon finally did.
In my view, training is the limit that will decide how this plays out. Clinicians may save time yet lack clear instruction on model limits, when to override, and how to recognize errors. Rules for buying, credentialing and liability will need to state who is responsible when guidance is followed and when it is ignored. The tools will improve through step-by-step use in clinics. Safety will depend on whether oversight keeps the same pace.
