When an artificial intelligence system makes a mistake in a hospital, who pays? The American Medical Association answered that question on October 29, 2026, with a new clinical AI liability framework that divides responsibility among the people and organisations best placed to prevent the harm. Under the AMA clinical AI liability framework, developers answer for what they build, hospitals answer for how they choose and monitor the tools they deploy, and doctors answer only for the clinical judgment they can meaningfully exercise.
The AMA announced the framework from its Chicago headquarters on the morning of October 29, describing it as the first contribution in a new line of work on one of the most urgent unresolved questions in healthcare AI: who is responsible when something goes wrong. The association said existing liability frameworks were not designed for technologies in which physicians, developers and healthcare organisations each exercise different levels of control over the same system.
That mismatch, the AMA argued, creates two problems at once. Physicians may face liability for risks they cannot reasonably prevent, because they never see the data, the model or the reasoning behind an AI output. At the same time, patients may struggle to seek remedies when AI causes harm, because it is unclear which party's failure actually caused it.
AMA President Willie Underwood III, MD, MSc, MPH, said: “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. As AI assumes more clinical functions, developers, healthcare organizations and physicians each have important responsibilities. We need clear expectations that protect patients, support sound clinical judgment, and ensure accountability remains with those best positioned to prevent harm.”
The framework recognises that more than one party may bear liability for a single incident, with accountability aligned with each party's ability to prevent the harm. It also stresses that physician oversight cannot substitute for adequate evidence, appropriate system design, or the conditions needed to make clinical review meaningful — in other words, a doctor cannot be expected to catch an error the system never showed them.
The liability brief follows an earlier AMA governance framework for agentic AI, released on October 16, 2026, which set out how increasingly capable AI systems should be allowed to take on consequential clinical functions, under what conditions, and with whose authority. The liability framework is the first step in delivering on the commitment made then to advance the harder questions of liability, licensure and regulation.
The timing matters because AI is already woven into everyday clinical work in the United States. Algorithms help read scans, triage patients, draft notes, flag deterioration and prioritise waiting lists. In many of those workflows, the physician sees only an output — a reordered patient list, a risk score, a suggested diagnosis — not the reasoning behind it. A 2026 AMA physician survey on augmented intelligence found that more than one-third of respondents chose clear liability frameworks as the regulatory action most likely to raise their trust in AI, a sign of how much the uncertainty weighs on adoption.
The legal backdrop is fragmented. There is no single federal rule assigning liability for clinical AI. States have begun to fill the gap in different ways: California's AB 3030 requires disclosure for generative AI communications, and Texas mandates written disclosure before an AI system is used in treatment. Court cases are also testing the boundaries — including a lawsuit filed in October 2026 by the parents of a young woman with epilepsy who died after, they allege, a chatbot coached her to taper off her anti-seizure medication, a case US News Pulse covered in full. Against that patchwork, the AMA's framework is an attempt to set a principled baseline before courts and legislatures set one piecemeal.
The framework is, however, early analysis from the AMA Center for Digital Health and AI. It adopts no new policy and creates no new legal requirement. Its influence will depend on whether lawmakers, courts, hospitals and vendors take up its division of responsibility.
The heart of the framework is a simple principle: accountability should follow control. Each party bears the risks it is best placed to foresee and prevent, and no party should be made to carry risks that sit entirely in someone else's hands.
Developers and manufacturers, the framework says, should demonstrate and maintain the safety and effectiveness of their products and address risks within their control. That includes evidence that a system works as intended, in the populations and settings where it will actually be used, and a duty to keep monitoring and fixing it after deployment rather than treating launch as the end of responsibility.
Healthcare organisations and other deployers should be accountable for how AI systems are selected, implemented, governed and monitored in real-world settings. In practice, the accompanying issue brief says, that means hospitals, health systems and practices should maintain processes to choose each tool carefully, validate it locally where it will run, monitor its performance in use, and pull it when warranted. Vendor contracts, the brief adds, should not shift responsibility for defects within a developer's control onto physicians or deployers.
Physicians should be responsible for 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. The brief gives a concrete example: a physician should not be held liable for an AI error they cannot reasonably detect, such as a triage model that quietly deprioritises a case while the doctor sees only a reordered patient list. Physicians should also be free to decline unproven AI without penalty, so that refusing a tool they do not trust is not treated as a failure of care.
Patients, finally, must retain meaningful avenues for seeking accountability and appropriate remedies when harm results from defective AI systems or other actionable failures. The framework treats patient recourse not as an afterthought but as one of the four pillars: if no one can be held to account, the system has failed, however advanced the technology.
Jennifer Goldsack, senior vice president and inaugural leader of the AMA Center for Digital Health and AI, framed the brief as the start of a longer programme of work. In a statement accompanying the release, she said patients must have meaningful avenues for recovery when AI causes harm, physicians need the evidence, information and support to exercise clinical judgment with confidence, and developers and health systems need clear expectations about their responsibilities so they can innovate and implement AI responsibly.
She added that the brief is a starting point and that the AMA is already advancing the next steps it lays out, working alongside physicians, patients, legal experts, policymakers and technology leaders. The association said it will bring those groups together to shape responsible governance, including on the licensing and regulation questions that follow from liability.
Reaction from the wider health-technology world was still forming as the framework was published, but the direction of travel is clear from the survey data the AMA itself cites: physicians want clarity before they trust AI more deeply, and more than a third named liability as the single most trust-building regulatory step. Health-system leaders, for their part, face a new expectation that their duty of care covers the selection, validation and monitoring of clinical AI — not just the purchase of it.
The framework also lands in a week when AI safety for vulnerable users is firmly in the spotlight. New York's new law, signed on October 28, bans AI chatbots from pretending to be human or forming personal relationships with minors, and the AMA's brief covers the other end of the risk spectrum: the clinical setting, where an error can change a diagnosis, a treatment or a life.
Liability is not an abstract legal question. It decides who invests in safety, who discloses a failure, and who a harmed patient can actually sue. When liability is unclear, the default in practice is often that the doctor — the most visible human in the chain — absorbs the blame, even when the fault lay in a model they did not build, data they did not see, or a deployment decision they did not make. That discourages doctors from using AI at all, or, worse, encourages them to follow AI outputs uncritically because second-guessing them feels riskier than obeying them.
By splitting responsibility according to control, the AMA framework aims to break that trap. Developers have a reason to prove safety and keep proving it. Hospitals have a reason to validate tools locally instead of assuming a vendor's claims transfer to their patients. Doctors have a reason to exercise real judgment — and to say no to a tool they cannot trust — without fear that refusal itself becomes the liability. And patients keep a route to redress when the system fails them.
The framework also matters for the pace of AI adoption in American healthcare. Hospitals are being offered an expanding menu of AI products, from ambient documentation to diagnostic support, often with limited evidence about how they perform outside the vendor's test environment. A clear expectation that deployers must validate and monitor locally could slow some purchases — but it could also make the purchases that do happen safer and more defensible, which is exactly what a cautious health system needs before it lets AI near the bedside.
For patients, the most practical consequence may be transparency. The framework's logic points toward a world in which patients can find out when AI contributed to a decision about their care, what its limitations were, and who was responsible for checking it. No federal requirement currently guarantees that disclosure, and the AMA's brief does not create one — but it gives patient advocates a principled argument for one.
The AMA says this brief is the first contribution in its liability work, not the last, and that it is already advancing the next steps alongside physicians, patients, legal experts, policymakers and technology leaders. Watch for three developments. First, whether state legislatures or Congress pick up the framework's division of responsibility in bills on AI in healthcare. Second, whether courts cite it when apportioning fault in the growing number of AI-related malpractice and product-liability cases. Third, whether hospitals rewrite vendor contracts to stop responsibility for developer-controlled defects being shifted onto clinicians — the practice the brief explicitly calls out.
In the shorter term, health systems that deploy clinical AI will face questions from their own clinicians and risk managers: which tools have been validated locally, how performance is being monitored, and what happens when a tool is pulled. Those questions are now, in the AMA's framing, part of a health system's duty of care, not an optional extra.
US News Pulse will follow how the framework is received by hospitals, developers and regulators, and whether it becomes the template for the first real clinical AI liability rules in the United States.
It is a framework released by the AMA Center for Digital Health and AI on October 29, 2026, titled 'Liability for Clinical AI', which aligns liability for harm from clinical AI with each party's authority and ability to prevent that harm.
Under the framework, responsibility is shared according to control: developers for the safety and effectiveness of what they build, healthcare organisations for how they select, implement, govern and monitor AI, and physicians for the clinical judgment they can meaningfully exercise. Patients must keep avenues for accountability and remedies.
No, under the AMA's approach. Physicians should not be held liable for AI errors they cannot reasonably detect, such as a triage model that deprioritises a case while the doctor sees only a reordered patient list, and they should be free to decline unproven AI without penalty.
No. It is early analysis from the AMA Center for Digital Health and AI that adopts no new policy and creates no legal requirement. Its influence depends on lawmakers, courts, hospitals and vendors adopting its approach.
Because AI is already taking on consequential clinical functions while existing liability frameworks were not designed for systems in which physicians, developers and healthcare organisations each control different parts. More than a third of physicians in the AMA's 2026 survey said clear liability frameworks would most raise their trust in AI.
Related coverage: OpenAI lawsuit: parents say ChatGPT coached daughter off seizure drug, New York AI chatbot law: Hochul bans bots pretending to be human for kids and No AI bailout: House Democrats tell Congress not to rescue failed AI firms.
Leave a Comment
| Today's Date:
