AI in the operating room: where it helps, and where it doesn’t

I have been asked, in every professional forum I have attended for the last two years, some version of the same question: is artificial intelligence going to replace surgeons? The question is the wrong one. The right question — the one that actually determines whether patients benefit — is more modest and more difficult. Where, specifically, does AI help a surgeon do a better job today? And where does it not?

Where it helps

Pre-operative risk stratification. Machine learning models trained on large surgical databases — NSQIP, MBSAQIP, institutional datasets — are now measurably better than the previous generation of risk calculators at predicting 30-day complications. The gain is small but real. Used well, they help identify the patient who needs an extra week of prehabilitation, or a different anesthetic plan, or a frank conversation about whether surgery is the right choice at all.

Intra-operative video analysis, in review. Surgical video is a rich data source that we have historically discarded. AI systems can now segment instruments, identify anatomic landmarks, and flag deviations from a critical view. Used after the fact — in M&M conferences, in fellowship review, in personal quality improvement — this is transformative. A fellow can watch their own case with the safety-of-critical-view score overlaid. That kind of concrete, case-specific feedback did not exist five years ago.

Post-operative surveillance. Continuous vital-sign analysis on the ward can detect the physiologic signature of a leak or a bleed hours before the bedside nurse or the on-call resident notices it. The published data on this is incomplete, but the direction is clear. Early warning is the highest-leverage moment in post-operative care, and the machines are getting good at it.

Documentation. Not glamorous, but real. Ambient scribes and structured note generation give a surgeon back an hour or two a day. That hour goes to patients, or to teaching, or to sleep. All three matter.

Where it doesn’t help — yet

Real-time intra-operative guidance. The demonstrations are impressive. The clinical evidence is thin. Systems that promise to identify structures you cannot see, or to warn you before a dangerous move, are almost all in the research-prototype stage. Deploying them in a live case, in the absence of prospective trials, is not innovation. It is experimentation on patients who did not consent to be experimented on.

Autonomous surgical action. The trials of autonomous suturing and autonomous soft-tissue tasks are scientifically interesting and, at the pace of current progress, not clinically relevant for at least a decade. The engineering problems that remain — tissue heterogeneity, unpredictable bleeding, the management of the unexpected — are exactly the problems a trained surgeon exists to handle.

Diagnosis from images without a physician in the loop. The FDA has cleared several standalone imaging AIs. The published error modes are consistent: the systems are excellent on the distribution they were trained on and unreliable on anything outside it. Every one of them still needs a physician reading the study, because the physician is the one who knows what to do when the model is wrong.

What surgeons owe the technology

Skepticism, and engagement. Both. It is not acceptable, at this point in the field, for a surgeon to dismiss AI as a fad and refuse to learn what it does. Nor is it acceptable to adopt every product a vendor pushes into the operating room because it has a neural network inside it. The obligation is to understand the tool at the level a surgeon understands any other tool: what it does, what it does not do, how it fails, and how a failure would harm the patient.

I teach residents that the correct question about any AI system in surgery is the same question you would ask about a new stapler. Where does it help? Where does it hurt? What does the evidence say? Who is accountable when it fails? If a surgeon cannot answer those four questions about a system they are about to use in a case, they should not be using it in the case.

The realistic future

In ten years, AI will be an unremarkable part of surgical practice, the way intra-operative imaging or electronic health records are today — useful, sometimes annoying, occasionally decisive, largely invisible. The surgeons who prepare now will use those systems well. The surgeons who wait for the technology to be settled will find themselves practicing to a lower standard than their patients deserve. That has always been the pattern with new technology in surgery, and there is no reason it will be different this time.