Should liability for an AI-influenced misdiagnosis sit with the clinician, the hospital, or the model vendor?

The shift

Diagnostic synthesis — reading a full record against the literature and returning a ranked differential — goes from scarce specialist cognition to abundant, near-free output that now sits inside the clinical workflow. So AI materially shapes the diagnosis, but the liability framework it enters still assumes one licensed human formed that judgment alone. The influence moved; the accountability model didn't.

The axioms

• Liability follows the mind that made the decision — whoever reasoned to the diagnosis owns the harm when it's wrong. Rests on the scarcity of the reasoning act: one person did the cognitive work, so one person answers for it.
• Every diagnosis has a single licensed owner who is answerable. Rests on accountability being non-transferable and attached to a license.
• The standard of care is what a reasonable clinician would have done — measured against peer practice. Rests on there being a stable, slow-moving consensus of what "reasonable" means.
• A tool's maker is liable for the tool's defects, not for the practitioner's use of it (the "learned intermediary" logic). Rests on a clean line between a product's design and a professional's judgment in applying it.
• The patient harmed by a wrong call has a defined party to sue and recover from. Rests on there being a solvent, identifiable, accountable defendant.
• A clinician is accountable for the tools they choose to rely on, and is presumed able to evaluate them. Rests on tools being interrogable — you can, in principle, understand why the instrument said what it said.

Invalid axioms

1. Liability follows the single mind that made the decision. Once a model produces the differential and the clinician accepts, edits, or overrides it, the "mind" is distributed across a vendor's training choices, a hospital's deployment settings, and a clinician's judgment about when to trust it. The reasoning act is no longer scarce or singular. The habit-trap: malpractice suits, insurance, and hospital risk models still name one defendant and reconstruct one chain of reasoning, as if a single person did all the thinking — so they litigate a fiction and misprice the actual risk.

Unchanged axioms

1. A licensed human must own the clinical decision and be answerable for it. This is a legal and regulatory fact, not a capability gap: the FDA clears AI as decision support, licensure sits with people and institutions, and a model can't hold a license, carry insurance, or be sanctioned. Better models don't move this — the more capable the tool, the more the "human in the loop" requirement is doing the actual work of locating accountability.
2. The standard of care still governs — someone is measured against what a reasonable clinician would have done. The content of that standard is moving (see NEW), but the mechanism doesn't disappear. There is still a benchmark of reasonable practice, and deviation from it is still what liability turns on.
3. The patient still needs a defined, solvent party to recover from. Recourse is the point of the whole system. Distributing the decision across three parties doesn't reduce the patient's claim to being made whole; it only makes the defendant harder to name. The need is untouched; satisfying it got harder.
4. Judgment about when to trust or override the tool stays with the accountable human. The scarce act isn't producing the read — it's deciding, on a specific patient with specific stakes, whether this is a case where the model's confidence is earned. That call sits with the person who answers for it.

New axioms

1. The liability chain across clinician, hospital, and vendor is unsettled, and each party has an incentive to point at the others. The vendor disclaims via terms of use and "decision support only" labeling; the hospital points to the clinician's independent judgment; the clinician points to a tool they were encouraged or required to use. Absent a rule that allocates the loss, the patient's recourse degrades into a multi-party fight — and the deepest-pocketed, most-disclaimed party (the vendor) is often the best-insulated. This hinges on fast-moving law: product-liability doctrine, FDA framing, and the first appellate rulings on AI-influenced error are actively forming as of mid-2026, and a handful of decisions or a statute could reset the whole allocation.
2. The standard of care is shifting toward "should have used the AI" and "should have overridden it" — in both directions at once. Once a tool is widely adopted and demonstrably catches misses, not consulting it can become the deviation; but deferring to it when a reasonable clinician would have caught its error is also a deviation. The clinician is exposed on both sides of a line that peer practice hasn't yet drawn. This is the fastest-moving call here: the standard follows adoption, and adoption is climbing quarter over quarter.
3. Clinicians are being held accountable for tools they can't fully interrogate. The old presumption — that a professional can evaluate the instrument they rely on — breaks against models whose reasoning is opaque, whose training data is proprietary, and whose failure modes are probabilistic rather than legible. "You're responsible for how you used it" assumes you could understand it well enough to use it responsibly. Nothing yet defines what responsible reliance on a black box even looks like.
4. Vendors are disclaiming liability faster than the law is assigning it. "Decision support, not a medical device decision" and contractual liability caps let vendors capture the upside of influence while offloading the downside onto the licensed human. Whether product-liability and "defective design" doctrine can reach probabilistic-but-not-defective model output is unresolved — and until it is, the party shaping the diagnosis is the party least likely to pay for it.

Where it breaks

"Liability follows the single mind" (invalid) collides with "the standard of care is shifting toward should-have-used / should-have-overridden" (new): the system still reconstructs one accountable reasoner after the fact, but it now judges that reasoner against a standard that presumes an AI was in the loop. The clinician is simultaneously treated as the sole author of the decision and faulted for how they interacted with a tool they didn't build and can't fully inspect — sole authorship for the blame, shared authorship for none of the control.

A second collision: "the patient needs a solvent, identifiable defendant" (still holds) meets "vendors are disclaiming faster than the law assigns" (new). The party with the deepest influence on the diagnosis and the deepest pockets is contractually the hardest to reach, so the recourse the system exists to provide routes, by default, to the individual clinician — the least-resourced link, and the one with the least control over the tool's behavior. Nobody has decided whether that's acceptable; it's just where the disclaimers currently push the loss.

ai-axioms.vercel.app/axioms/healthcare/medicine-malpractice-ai-axioms