Will AI Replace Surgeons? A Realistic Look

Will AI replace surgeons? No, not on any horizon a current trainee needs to plan for. It will change which parts of the job take your time, and it will make one skill more valuable than it has been: judging when a machine's output is wrong.

This is an opinion piece. The evidence on what orthopaedic AI can and cannot do (fracture detection, outcome prediction, planning software, regulation) is set out in our evidence review of artificial intelligence in orthopaedic surgery. What follows is the argument we draw from it.

What the software actually does

Strip away the marketing and almost every clinical AI product does one narrow thing. It looks at one kind of image and marks one kind of finding, or it takes a table of variables and returns a probability. That is useful. A wrist radiograph read at 3 a.m. by a tired junior is read better with a second pair of eyes, even a synthetic one.

But notice how these tools are framed by the people who regulate them. The US device category for fracture-detection software ends its definition with the sentence: "Primary diagnostic and patient management decisions are made by the clinical user." The fracture-detection products that have reached the market were cleared as aids to a clinician, not as replacements for one, and the best-supported use, as a second reader, depends on the clinician staying in the loop.

The jump from "spots a fracture" to "manages the patient" is not a matter of more data. Management means knowing that this patient lives alone, has a carer's role, smokes and would rather keep a stiff wrist than have another operation. None of that is on the radiograph.

Why the operation itself resists automation

Pattern recognition works best when the patterns are stable. Operating is the opposite. Anatomy in textbooks is an average; anatomy on the table has been altered by previous surgery, scar, infection, tumour or simple variation. The surgeon reads tissue by feel as much as by sight: whether fascia will hold a suture, whether bone will hold a screw, whether a plane is real or about to tear.

The orthopaedic robots in use today do not change this picture. They help a surgeon carry out a plan the surgeon has made and approved, and the surgeon is in the theatre throughout. Our review of robotic arthroplasty sets out what they have and have not been shown to improve. The hard part of surgery is not the cut that was planned. It is the moment the plan stops fitting the patient: unexpected bleeding, a fracture you did not cause on purpose, an implant that will not seat, an anaesthetic emergency. Those moments need improvisation drawn from experience of many unlike cases, which is exactly where systems trained on past examples are weakest.

The question of who carries the risk

Replacement would require someone else to carry the risk. At present nobody is offering to.

In the UK, NHS England's guidance on AI scribing products requires a clinician to review every output and warns that NHS organisations may remain liable where a non-delegable duty of care is involved. The device classifications say the same thing in regulatory language. Whatever the software suggests, the decision is yours, and so is the consequence.

That creates a specific hazard: automation bias, the tendency to accept a confident machine output over your own conflicting judgement. "The software said there was no fracture" will not defend a missed scaphoid fracture in a patient who was tender in the snuffbox and over the tubercle. If an AI-generated plan looks wrong for the patient in front of you, you must be prepared to discard it. That takes knowledge, and it takes the confidence to disagree with a screen in front of a team.

How the job changes

If AI does not replace surgeons, it does move work around.

Paperwork shrinks, a little. Ambient scribes that draft the clinic letter are the most visible change. A randomised trial published in 2025 found a modest effect: one product cut time spent writing notes by about a tenth and the other made no significant difference. That is worth having, but it is not the hours-a-day transformation in some vendor claims. The letter still needs reading before you sign it.

Teams get more complex. Imaging software, planning platforms, navigation and monitoring each add a stream of information, and someone has to reconcile them. The surgeon's role shifts towards that of a team leader who weighs inputs, including from machines, and is explicit about which one is being overruled and why. Theatres already work best when anyone can speak up; the same principle has to apply to the software.

Planning becomes reviewing. Where segmentation and templating are automated, the surgeon's job becomes checking a proposed plan rather than drawing it. That is faster, and it is only safe if you could still have drawn it yourself.

What this means for training and exams

The fundamentals do not change. Anatomy, pathology, biomechanics, the principles of fixation and the management of complications remain what you are examined on, and what you rely on when a plan fails. A trainee who lets software do the templating for five years will not know when the template is wrong.

What is added is appraisal. It is reasonable to expect more questions, at the bedside and in vivas, of the type "the software says X; what do you do?" The answer has the same shape as any good viva answer: clinical findings first, then the investigation, then a reasoned decision, with the tool treated as one more test with known strengths and failure modes. Being able to ask whether a model was externally validated, what population it was trained on and what its false-negative rate looks like is now part of critical appraisal. Our guide on how to read a research paper critically applies to AI studies with little modification.

What to do about it

Three things are worth doing now.

1. Keep your hands and your judgement in practice. Do the templating yourself before you look at the automated plan. Examine the patient before you read the AI report. Your own reading is the control against which the tool is measured.
2. Invest in the parts of the job no model is close to doing. Consent, breaking bad news, shared decisions with a frail patient and their family, leading a theatre team when a case goes wrong. These are skills, they can be practised, and they are not being automated.
3. Learn enough about the tools to argue with them. Know what your hospital's software is cleared for and what it was tested on. When it is wrong, report it. The clinicians who shape how AI is used will be those who engage with it critically, not those who ignore it or trust it blindly.

The surgeon of the next twenty years will do less transcription and more checking, will lead a team that includes software, and will still be the person who decides and answers for the decision. That is a change in the job, not the end of it.