AMA redraws liability lines for clinical AI
The American Medical Association is pressing for a different legal architecture around clinical AI: one that ties responsibility to who can actually prevent harm. The group’s newest framework argues that existing liability rules are too blunt for software that is designed, marketed, deployed, and supervised by different parties with different degrees of control. In practice, that leaves a dangerous gap. Physicians can end up exposed for risks they cannot reasonably avert, while patients may struggle to find a remedy when an AI-driven recommendation contributes to injury.
The framework is built around a simple premise with major medicolegal consequences: liability should follow meaningful control. When risk is shared, responsibility can be shared as well, but not in a way that makes every participant interchangeable. A physician, a vendor, and a hospital may all influence an AI-enabled workflow, yet each has a different role in the chain of decision-making, validation, deployment, and oversight. The AMA’s position is that current frameworks blur those distinctions instead of clarifying them.
Why existing liability rules do not fit clinical AI
Traditional liability models were developed for tools with clearer lines of authorship and use. A drug has a labeled indication, a device has a defined regulatory pathway, and a clinician usually has a recognizable duty of care at the bedside. Clinical AI disrupts that arrangement because the system may learn from data, adapt over time, and provide recommendations that depend on the context in which they are used. A disclaimer in a contract does not change the fact that a system can steer care, shape workflows, and alter outcomes.
AMA identifies several problem areas in the status quo. Product liability remains uncertain when the technology is not a fixed object but an evolving system. Developer language in contracts and disclaimers can distance responsibility from the actual risk control environment. Healthcare organizations, meanwhile, may deploy AI without a clear understanding of whether they are expected to perform local testing, monitor outcomes, or detect performance drift over time. Each of those failures can matter clinically long before a courtroom sorts out fault.
A short sentence captures the core risk: the patient still absorbs the harm.
That reality is what pushes the AMA toward a liability framework that is more granular than conventional medical malpractice rules. If an institution selects an unproven model, integrates it into clinical workflow, and fails to monitor its real-world performance, the burden should not land solely on the physician who is asked to trust the output. If a developer introduces a system without adequate validation or with known limitations that are not transparent to users, accountability cannot disappear behind a contract clause. The framework is trying to map legal responsibility onto the actual mechanics of harm.
Physician judgment remains central, but not limitless
For physicians, the AMA says responsibility should extend to the clinical judgment they can meaningfully exercise, based on the system’s intended use, its known limitations, and the information available at the point of care. That wording matters because it rejects the idea that clinicians should serve as a catchall backstop for every failure in the AI stack. A doctor can interpret a recommendation, question a suspicious output, and decline to follow a tool that does not fit the clinical situation. What a physician cannot do is make a poorly designed system safe simply by being present.
The framework is explicit that oversight from a physician cannot substitute for adequate evidence, appropriate system design, or a clinical environment that allows review to be meaningful. In bedside terms, that means physician sign-off is not a cure for weak validation, incomplete transparency, or workflows that leave no time to interrogate an output. The legal theory is not just about blame after an error; it is about placing obligations where prevention is actually possible.
Physicians should also have the authority to refuse an unproven AI tool without organizational penalty if clinical validation, usefulness, or its place in the standard of care remains uncertain. That protection is important because forced adoption can turn a clinician into the final checkpoint for technology that has not earned trust. Once a hospital mandates use, the system’s failure is no longer just a technical issue. It becomes a patient-safety issue with a medicolegal tail.
Developers, hospitals and patients all carry separate stakes
The AMA’s framework also creates room for patient recourse when AI causes harm. That is not a trivial point. If patients cannot identify who controlled a faulty recommendation, they may have no practical route to compensation or accountability even when the injury is traceable to the technology. A liability framework that leaves patients stranded is not just legally unsatisfying; it weakens confidence in digital medicine at the moment the field is trying to expand.
Developers, healthcare organizations, and physicians each occupy a different node in the care pathway. Developers influence the model’s architecture, training, testing, and disclosure of limitations. Healthcare organizations decide whether the system is validated locally, how it is integrated, and whether outcomes are monitored after deployment. Physicians interpret the tool in real time, using professional judgment in the context of patient history, symptoms, and available data. The AMA framework assigns liability in a way that reflects those separate functions rather than collapsing them into a single bucket.
Governance questions now extend beyond liability alone
This liability framework is one piece of a broader AMA effort to define how clinical AI should be governed. The organization recently laid out wider principles for AI governance, and those principles are meant to guide work on licensure, regulation, and other policy questions as the technology moves faster than the legal system. The same logic runs through both efforts: clinical authority should be earned through evidence, oversight should evolve with risk, and accountability should remain attached to the parties best positioned to prevent harm.
Willie Underwood III, M.D., framed the issue in direct patient-safety terms, arguing that patients must have meaningful avenues for recourse when clinical AI causes harm and physicians must be able to use these technologies with confidence that responsibility reflects what they can reasonably know and control. John Whyte, M.D., has said the profession cannot wait for these questions to be answered through trial and error in patient care. Those comments reflect the central concern behind the framework: AI is moving into clinical work now, and the liability structure around it cannot remain vague while deployment accelerates.
The framework does not settle every open question. It does, however, make one point unmistakable: responsibility in clinical AI should not vanish into disclaimers, and it should not be dumped onto the physician at the bedside when the real failure occurred upstream.