AMA Calls for US Laws on Clinical AI Liability

A clinical AI recommendation can pass from a software vendor to a hospital and then to a physician. The AMA wants lawmakers to consider what each party could have done to prevent an error.

The American Medical Association’s House of Delegates adopted new advocacy policies on clinical AI at its November Interim Meeting, which ended November 9. The AMA says it will seek federal and state legislation on liability for errors from clinical AI products and protections for physicians’ independent judgment. The vote sets the association’s advocacy position. The changes would require lawmakers to act. AMA meeting summary and Axios reporting on the policy.

What the AMA wants lawmakers to address

The AMA wants vendors, manufacturers, or distributors to bear responsibility for harm caused by an erroneous product output when the physician used the tool as intended and exercised appropriate clinical judgment. If a hospital, health system or payer mandates a particular tool, the AMA wants the organization that required it and the vendor to share responsibility for harm caused by an error originating in the mandated product, subject to the physician’s exercise of appropriate judgment.

The delegates also called for insurance for clinical AI vendors, disclosure of known limitations and performance across patient groups, and restrictions on contracts that shift responsibility for product failures onto the physician or health system using the tool. The AMA’s account of the adopted policies.

What a procurement team could document

The AMA’s October framework ties responsibility to the ability to prevent or reduce harm. A procurement team can use that idea to design a tabletop review. Start with a fictional case where the software makes a recommendation using outdated input. Then trace who could detect the missing or outdated information, who could change the deployment, and what the clinician would see. The AMA’s October framework.

Who can act	What to check in the simulated review
Vendor	Which model version ran, what inputs it expected, and which known limitations it disclosed.
Hospital or health system	Whether the software was required and whether someone could suspend it or switch to a fallback workflow.
Treating physician	What input evidence the interface showed and whether it allowed the physician to reject the recommendation.

One useful question is whether a clinician can see the timestamp of the data that led to the recommendation. If the interface hides it, a review record can show who saw the output while leaving a gap about what they could verify. Documenting that gap during a simulated review gives the vendor and hospital a concrete design issue to discuss. Keep the exercise focused on the evidence each team can see and the decisions it controls.