What changes for surgery with AI?

The shift

Everything upstream and alongside the cut goes abundant: pre-op planning, imaging segmentation, and intra-op guidance and decision-support move from scarce specialist time to fast, near-free, and generated on demand. The physical act itself — the accountable person cutting a specific body, and adapting when that body doesn't match the plan — does not.

The axioms

• Operating on a body requires trained hands performing physical action in the physical world.
• A named, licensed surgeon must own an irreversible outcome and answer for it.
• Judgment when the body surprises the plan — anatomy that doesn't match the scan, bleeding that wasn't predicted — is scarce and hard-won.
• Pre-op planning (mapping the approach, sizing the resection, anticipating the anatomy) is gated by scarce expert cognition and time.
• Reading imaging and localizing the target is gated by scarce radiology/surgical expertise.
• Intra-op guidance and decision-support are limited by what a single team can hold in its head under time pressure.
• Surgical skill is built by accumulating manual repetitions under supervision.
• Robotic systems extend a surgeon's hands but are directed move-by-move by that surgeon.
• Patient trust in a surgeon is built through relationship and the standing to be accountable.

Invalid axioms

1. Pre-op planning is gated by scarce expert cognition and time. A model can segment the anatomy, propose an approach, size the resection, flag likely complications, and simulate variants in minutes rather than hours of a senior surgeon's evening. The habit-trap: planning time is still booked, staffed, and reimbursed as scarce senior-clinician labor, and complex-case conferences still run as if generating the plan were the expensive step rather than checking it.
2. Reading imaging and localizing the target is gated by scarce radiology/surgical expertise. Segmentation, 3D reconstruction, and target localization are now abundant and instant. The habit-trap: workflows still route every read through scarce human interpretation as the primary source rather than as verification of a machine read.
3. Intra-op guidance is limited by what one team can hold in its head under time pressure. Overlays, live structure identification, and decision-support can now surface the relevant reference and prior cases continuously during the operation. The habit-trap: the OR is still designed around the unaided memory and attention of the team, with no defined role for who owns, checks, or overrides a live AI read mid-procedure.

Unchanged axioms

1. The physical act of operating requires trained hands and physical presence. Cutting, dissecting, suturing, and controlling bleeding in a specific living body are not token-generation problems. Robotic assist improves precision, tremor filtering, and access in narrow fields, and will keep improving — but through mid-2026 it is teleoperated, surgeon-directed move-by-move, not autonomous. Autonomous execution of a full open or complex procedure is not close; supervised autonomy for narrow, well-bounded sub-tasks (e.g. specific soft-tissue steps demonstrated in the lab) is the live frontier and the part of this call most worth re-checking as the robotics trajectory moves.
2. A named, licensed surgeon must own the irreversible outcome. Liability doesn't transfer to a model or a robot vendor. When a resection takes the wrong margin or a vessel is nicked, someone with a license and a name answers for it. This is a legal and institutional fact, not a capability gap, so better models don't move it.
3. Judgment when the body surprises the plan stays scarce. The value concentrates exactly where the pre-op plan and the imaging were wrong — unexpected adhesions, aberrant anatomy, a bleed that changes the whole operation. These are the low-pattern, high-stakes moments where matching against "everything ever recorded" is weakest and a hallucinated read is most dangerous, and where an accountable human has to decide in seconds with the patient open.
4. Being confidently wrong is unusually dangerous here, and irreversible. A plausible-but-wrong intra-op overlay — a mislabeled duct, a misidentified plane — can cause direct, permanent harm with no redo. This raises the bar on verifying AI guidance rather than lowering it, and the verification has to happen live, hands occupied, under time pressure.
5. Patient trust and the standing to be accountable stay human. Consent for an irreversible procedure, and the relationship a patient enters when they let someone operate on them, run on trust in an accountable person, not on the quality of a generated plan.

New axioms

1. Accountability when AI-guided or robotic assist is involved in a bad outcome. Malpractice and product-liability frameworks assume the operating mind and the accountable party are the same person. When the plan was AI-generated, the overlay was AI-driven, and the instrument was robotic, the chain of responsibility across surgeon, hospital, and vendor isn't settled — and it needs to be resolved before, not after, the adverse event.
2. Automation bias in the OR. When a live overlay confidently labels a structure, the pull is to trust it precisely when the case has gone off-pattern — the moment it's most likely to be wrong. The problem is designing guidance the surgeon reflexively checks rather than defers to, under exactly the conditions that erode scrutiny.
3. The training pathway when trainees do fewer manual reps. Surgical judgment has been built by accumulating supervised repetitions — including the slow, effortful planning and the difficult intra-op calls. If planning is auto-generated and more of the case is machine-guided, the pipeline that produces surgeons who can take over when the AI is wrong is untested at scale. This compounds: the scarce judgment in the STILL HOLDS bucket is produced by exactly the reps that are now easiest to skip.
4. Verifying AI guidance mid-operation. A plausible intra-op read has to be checked against ground truth while the surgeon's hands are occupied and the clock is running. There's no established protocol, role, or interface for auditing a live machine read without breaking the flow of the operation — verification is assumed to be free and instantaneous, which in the OR it is not.

Where it breaks

"Pre-op planning is gated by scarce expert cognition" (invalid) collides with "judgment when the body surprises the plan stays scarce" (still holds): the plan is now cheap to generate, but its value was never the plan — it was the surgeon internalizing the anatomy well enough to adapt when reality diverges from it. Auto-generating the plan removes the very rehearsal that built the ability to depart from it, and no one has resourced that as its own step.

A second collision: "intra-op guidance is limited by what one team can hold in its head" (invalid) meets "verifying AI guidance mid-operation" and "automation bias in the OR" (new). Continuous live guidance is being added to the operating field faster than anyone has defined who checks it, when, and with what authority to override — which moves the risk into the moment of the cut rather than removing it.

ai-axioms.vercel.app/axioms/healthcare/surgery-ai-axioms