Estimated reading time: 6 minutes · Last updated:
The American Medical Association has set out a clinical AI liability framework that splits responsibility for AI-related harm among developers, healthcare organizations and physicians. It sizes each party's share by that party's power to prevent the harm. Under the AMA's approach, doctors answer for judgment they can genuinely exercise at the bedside, not for risks they had no reasonable way to foresee or control. The framework comes from the association's Center for Digital Health and AI. It also says physicians should be able to refuse unproven tools without being penalised by their employers. It adds detail to a broader governance framework the AMA released earlier the same month, and it targets questions the group says existing liability law leaves unanswered.
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.
Willie Underwood III, President, American Medical Association
Key takeaways
- Control sets the share: The American Medical Association's clinical AI liability framework assigns accountability to developers, healthcare organizations and physicians according to each party's ability to prevent harm.
- No passing the buck: The AMA says that when control over an AI system is shared, responsibility is shared too, but one party's duties never cancel out another's.
- Right to refuse: The AMA says physicians should be able to decline AI tools whose clinical validation, usefulness or place in the standard of care is uncertain, without penalty from their organization.
- Gaps in current law: The AMA's Center for Digital Health and AI points to unclear product liability rules, developer disclaimers and undefined hospital duties such as local testing as gaps in today's liability frameworks.
- Part of a wider plan: AMA President Willie Underwood III and CEO John Whyte present the liability brief as one piece of a broader AMA clinical AI governance framework covering licensure and regulation.
Table of contents
- Key takeaways
- Why the AMA Says Current Liability Rules Fall Short
- How the AMA Ties Accountability to Control
- What Doctors Answer For, and What They Can Refuse
- Where the Liability Brief Fits in the AMA's Wider Plan
- Will the AMA's liability model take hold?
- What to be careful about
- Frequently asked questions
Why the AMA Says Current Liability Rules Fall Short
The AMA argues that today's liability structures were not built for software that takes on clinical tasks. In its view, they could expose physicians to "liability for risks they cannot reasonably prevent, while patients may struggle to seek remedies when AI causes harm." That is a failure on two sides at once. The doctor carries too much risk, and the injured patient may have nowhere obvious to turn.
The association lists three further gaps. The first is uncertainty over when product liability covers an AI system at all. Product liability is the branch of law that holds makers accountable for defective goods. The second is the way contract terms and disclaimers written by developers can separate who is legally exposed from who actually controls the risk.
The third gap concerns the hospitals and health systems that deploy these tools. The AMA asks whether they carry duties particular to AI. Examples include testing a system locally before use, or tracking patient outcomes once it is live. Existing frameworks, the group says, leave that unresolved.
How the AMA Ties Accountability to Control
The core idea is easy to state. The framework "recognizes that multiple parties may bear liability for the use of AI when something goes wrong," the AMA wrote. It then assigns accountability according to how well placed each party is to stop the harm. Whoever holds the levers carries the weight.
Where control is shared, so is responsibility. The AMA draws a firm line, however: shared duties are not swappable. One party bearing part of the blame does not wipe out what another party owes. A developer, a health system and a physician can each be answerable for the same incident, in different ways.
This principle bears most directly on the contract problem the AMA raised. Suppose accountability follows control. Then a disclaimer that pushes risk downstream sits awkwardly with a rule that keeps each party answerable for what it can manage. The AMA has not said how courts or regulators would apply that principle, and enforcement is the open question.
What Doctors Answer For, and What They Can Refuse
For physicians, the framework says doctors should "be responsible for the clinical judgment they can meaningfully exercise, based on the system's intended use, known limitations and the information available at the point of care." Responsibility, in other words, tracks what a clinician could reasonably know when the decision was made.
The AMA also rejects the idea that putting a human in the loop fixes a weak product. Physician review, it said, "cannot substitute for adequate evidence, appropriate system design or the conditions necessary to make clinical review meaningful." A doctor signing off on an untested tool is not a safeguard.
The most practical provision concerns refusal. The AMA says physicians should keep the authority to turn down an AI system that is unproven, or one "whose clinical validation, usefulness or place in the standard of care remains uncertain." Their organization, it says, should not penalise them for doing so. The standard of care is the level of treatment a competent clinician would be expected to provide in similar circumstances.
Where the Liability Brief Fits in the AMA's Wider Plan
The liability brief zooms in on one part of a broader clinical AI governance framework. The AMA and its Center for Digital Health and AI unveiled that framework earlier the same month. Its principles will shape how the association works with providers and other stakeholders on licensure and regulation.
AMA President Willie Underwood III set out the stakes in a release. "As AI assumes more clinical functions, developers, healthcare organizations and physicians each have important responsibilities," he said. He argued that accountability should rest with whoever is best placed to prevent harm.
AMA CEO John Whyte said these questions could not be left to be settled by trial and error in patient care, as first reported by Fierce Healthcare. "The AMA's new framework sets out our ambitious agenda to address these questions," he said. He added that the framework would guide the group's work with physicians, patients, technology developers and policymakers.
| Party | Problem the AMA identifies | AMA position |
|---|---|---|
| Developers | Contract terms and disclaimers can separate risk from responsibility | Accountability follows ability to prevent harm |
| Healthcare organizations | Unclear AI-specific duties such as local testing and outcomes monitoring | Should not penalise physicians who decline unproven tools |
| Physicians | May face liability for risks they cannot reasonably prevent | Responsible for judgment they can meaningfully exercise |
| Patients | May struggle to seek remedies when AI causes harm | Must have meaningful avenues for recourse |
Will the AMA's liability model take hold?
The case for
- Tying blame to control gives developers and health systems a reason to validate tools rather than rely on physician review.
- The right to decline unproven systems offers physicians a stated defence against workplace pressure to adopt untested tools.
- The brief sits inside a wider AMA governance framework that will shape the group's work on licensure and regulation.
The case against
- The framework is a position from a physicians' group, not law, so its effect depends on courts, regulators and contracts.
- The AMA has not said how a deploying organization's duty to test or monitor AI would be defined or checked.
- Developer contract terms and disclaimers that shift risk remain in place unless parties agree to change them.
What to be careful about
- Physicians could still face liability for AI-driven errors if courts do not adopt the AMA's control-based approach.
- Patients harmed by clinical AI may continue to lack a clear route to compensation while product liability coverage of AI stays ambiguous.
- Health systems that deploy AI without local testing or outcomes monitoring may face uncertain obligations until those duties are defined.
The bottom line
The AMA's framework does not change the law. It is a physicians' group stating how responsibility should be divided, and its force depends on whether hospitals, developers, regulators and courts accept its logic. Its strongest move is tying blame to control. That pushes back against contract language that leaves clinicians exposed for systems they did not design. Its most immediate effect could be the right to refuse: a doctor who declines an unvalidated tool now has the AMA's stated backing against workplace penalties. Enforcement remains the open question. The AMA has not explained how a deploying organization's duty to test or monitor would be defined or checked. Until that is settled, patients harmed by AI still face the gap the association describes.
What to watch
- Watch for the AMA to turn these liability principles into specific positions on licensure and regulation; no date has been set.
- Watch for whether hospitals and health systems formally take on AI duties such as local testing and outcomes monitoring; no timetable has been announced.
- Watch for any change in developer contract terms and disclaimers in response to the AMA's control-based model; no date has been given.
Frequently asked questions
Is the AMA clinical AI liability framework legally binding?
No. The American Medical Association has published a policy position on how responsibility should be divided. Whether it shapes outcomes depends on courts, regulators, contracts and the organizations that deploy AI.
Can doctors refuse to use AI tools at work?
The AMA says physicians should keep the authority to decline AI systems that are unproven, or whose validation, usefulness or place in the standard of care is uncertain. It says organizations should not penalise them for that choice.
What does the AMA framework mean for patients harmed by AI?
The AMA warns that under current frameworks, patients may struggle to obtain remedies when AI causes harm. AMA President Willie Underwood III said patients must have meaningful avenues for recourse.
What responsibilities could hospitals have under the AMA's approach?
The AMA asks whether deploying healthcare organizations hold AI-specific duties, such as testing systems locally and monitoring outcomes. It says existing liability frameworks leave those duties undefined.
Related reading
