World’s First Live AI Brain Surgery: Why Keeping Surgeons in Control Is the Real Test

The operation was kept confidential until late August while the team monitored the patient’s recovery. The patient, 48-year-old Rhys Hibbert, had an 11mm benign pituitary tumour pressing against the optic nerves and dangerously close to the internal carotid arteries. The tumour was completely removed without vascular injury, and his peripheral vision returned immediately.

That is an impressive clinical result. It is not yet proof that machines can safely navigate the brain.

What it does establish is a new test for medical AI: whether software can interpret the operating field as it changes, rather than simply analyse images captured before the surgeon begins.

From scans to the surgical field

Until now, much of the practical use of AI in neurosurgery has been built around pre-operative MRI and CT scans, including segmentation, diagnosis and planning. Those systems can be highly sophisticated, but they operate on a relatively fixed picture of the patient.

Surgery is different.

In an endoscopic transsphenoidal operation, the surgeon reaches the pituitary through the nasal cavity and skull base. The tumour itself is concealed behind bone and membrane. Once instruments enter the surgical field, anatomy is no longer simply a three-dimensional structure waiting to be identified. Tissue can move. Blood can obscure landmarks. Reflections, smoke and changes in lighting can alter what the camera sees.

A pre-operative scan cannot update itself when the operating environment changes.

The UCL system was designed around that problem. Developed at UCL’s Wellcome/EPSRC Centre for Interventional and Surgical Sciences and the Hawkes Institute, the software was trained using hundreds of annotated videos of pituitary operations. Clinical experts identified neural and vascular structures frame by frame, giving the model a much richer reference set than an individual surgeon could accumulate through personal experience.

During Hibbert’s operation, the software analysed the live endoscopic feed on a secondary monitor. It identified surgical instruments and critical anatomical structures, colour-coded them and generated dynamic “safe corridors” for the surgeon. The system also indicated the likely position of the optic nerves, carotid arteries and smaller vessels.

The important distinction is that it did not operate.

Professor Hani Marcus and Mr Danyal Khan retained control throughout the procedure. The AI was deliberately positioned as another expert offering guidance, not as an autonomous surgical agent.

That distinction will become increasingly important as these systems improve.

The experience gap is part of the problem

Pituitary surgery is already performed routinely. But routine does not mean uniform.

The briefing baseline puts the risk of incomplete tumour removal at 25 to 50 per cent, while the risk of damaging a major blood vessel is estimated at 0.5 to 2 per cent. A UK neurosurgeon performs only around 10 to 20 such operations a year on average.

An AI trained on hundreds of procedures approaches the problem from a different direction. It can retain patterns across a volume of surgical experience that no individual operator can personally accumulate.

That creates an intriguing possibility. Instead of replacing specialist judgement, AI could make some of the knowledge concentrated in major academic centres available elsewhere.

But this is where the first serious obstacle appears.

A model trained at UCL is not necessarily a model that will perform identically in another hospital. Different cameras, lighting conditions, instruments and operating techniques can alter the visual data reaching the algorithm. This is the familiar problem of domain shift, but in surgery it becomes more than a technical inconvenience.

A system that is slightly less accurate at a regional hospital is not merely producing a poorer image. It could be pointing a surgeon towards the wrong piece of anatomy.

Scaling the technology therefore means scaling the entire digital environment around it: computing capacity, compatible imaging equipment, data quality, monitoring and clinical expertise. The NHS may eventually have to decide whether AI-assisted surgery is an optional technological upgrade or part of the infrastructure required for high-risk procedures.

That could create a new inequality.

Hospitals unable to afford the necessary hardware and computing systems could find themselves operating under a different safety standard from well-funded academic centres.

The new surgical risk is the machine itself

There is an obvious temptation to describe the AI as an additional pair of eyes. But unlike a human colleague, an algorithm can fail in ways that are difficult to recognise from inside the operating theatre.

A model may misclassify a vessel because of blood, glare, or unusual tissue. A supposedly safe corridor could overlap with a structure hidden from the camera. A boundary might flicker between frames, distracting the surgeon precisely when concentration matters most.

These are not the textual “hallucinations” associated with large language models. They are spatial errors.

That makes human oversight essential, but human oversight has its own weakness: automation bias. Once a system has demonstrated strong performance, surgeons may begin to trust its recommendation even when their own clinical judgement points elsewhere.

The UCL operation was designed to preserve the opposite relationship. The overlay advises; the surgeon decides.

Whether that hierarchy can survive widespread adoption is a more difficult question.

There is also a less obvious danger. If surgeons increasingly rely on machine-generated spatial guidance, training could gradually change. Skills that are rarely exercised may weaken. Future specialists could become better at interpreting AI outputs while becoming less practiced at constructing the same anatomical map themselves.

That would be a poor trade if the system failed at the wrong moment.

Regulation is moving into unfamiliar territory

The regulatory problem is just as significant as the clinical one.

Medical regulators have traditionally assessed software according to defined functions and versions. Real-time intraoperative AI introduces something more fluid: a system continuously interpreting changing visual information while influencing decisions during an operation.

The European Union’s AI Act places relevant high-risk clinical AI within a framework of data governance, risk management, human oversight and post-market monitoring, alongside requirements arising from medical-device regulation. In the UK, the MHRA is developing its approach through its Software and AI Medical Devices Change Programme. The US FDA treats comparable systems as Software as a Medical Device, with mechanisms such as Predetermined Change Control Plans intended to govern certain future software modifications.

But the central problem is not simply whether an algorithm passes a regulatory test before deployment.

It is what happens when the algorithm changes, the hospital changes, the cameras change or the patient differs from the training data.

The legal question is harder still. If an AI system incorrectly identifies an artery and a surgeon follows its guidance, responsibility cannot easily be assigned to a single actor. The hospital owns the clinical environment. The surgeon controls the operation. The developer designed the model. The manufacturer may supply the hardware on which it runs.

Medical malpractice systems are not built around such a chain of shared responsibility.

The breakthrough is therefore also a warning

The funding model behind the UCL project points towards where this technology may go next. The system combines university research, NHS clinical expertise, public funding from the National Institute for Health and Care Research and investment from Google.

That combination matters because real-time surgical AI is not simply another medical application that can be installed as software. It requires clinical data, specialist annotation, computing infrastructure, compatible equipment, and years of validation.

The next stage will be larger, multi-centre clinical trials under NHS oversight. That is the point at which the technology faces its real test.

One successful operation can show that live AI guidance is possible. Multiple hospitals will have to show that it remains reliable when the lighting is different, the instruments are different, the surgeon is different, and the anatomy is less cooperative.

If it succeeds, the operating theatre will begin to look less like a place where a surgeon translates a static scan into physical action and more like a continuously updated computational environment.