AMA, AHA and Stanford Researchers Press FDA on Generative AI Device Oversight

The American Medical Association, the American Hospital Association and a group of Stanford researchers filed recommendations with the FDA on how it should regulate medical devices that use generative AI, as the agency's comment period on the topic closed on Monday, October 19.

The comments respond to a discussion paper the FDA released on August 18 through its Digital Health Center of Excellence. The paper proposes a two-axis framework for assessing risk, a premarket approach built on "competency assessment" that combines non-clinical benchmarking with clinical confirmation, and risk-proportionate postmarket monitoring. It also raises questions about foundation models and agentic AI systems. Comments were due October 19 under docket FDA-2026-N-7874.

Physicians: evidence first

The AMA said on Monday that it had sent a letter, dated October 16, to Rick Abramson, MD, director of the Digital Health Center of Excellence. According to the AMA's release, it urged the FDA to evaluate generative AI performance in real-world clinical settings, strengthen postmarket surveillance, involve physicians in defining the standard of care and keep manufacturers accountable for their products' performance. The AMA said oversight should reflect a device's intended use, potential for harm and performance across its lifecycle, and that it supports a least-burdensome, risk-based approach.

The association also offered to work with the FDA on best practices for clinician adjudication, ongoing performance review and identifying clinically consequential AI failures.

"Physicians have a critical role in defining what good performance looks like," AMA President Willie Underwood, MD, said in the release.

Hospitals: refine the risk test

The AHA said it supports a "flexible, risk-based regulatory framework" that protects patients without slowing responsible innovation. In its letter, the hospital group asked the FDA to refine its proposed risk heuristic, issue labeling guidance distinguishing "informational action directing" devices from "action taking" devices, and align premarket clearance with post-deployment evaluation and monitoring under a total product lifecycle approach. It also asked the agency to clarify that generative AI clinical decision support and administrative tools are not subject to FDA oversight.

Researchers: test what oversight really controls

A response from Stanford HAI and Stanford School of Medicine scholars, also dated October 19, made seven recommendations. They asked the FDA to require evidence that human oversight actually controls risk for the tasks a device performs, to define how the "consequences" axis of its framework is determined, and to strengthen benchmarking so it measures what it claims to measure. The group said real-world prospective testing should supplement or replace benchmarks, and recommended placing more monitoring responsibility on developers, combining automated and human monitoring, and requiring a prespecified postmarket monitoring plan.

Authors of the Stanford comment include Nigam Shah, Michelle Mello, Curtis Langlotz and Roxana Daneshjou.

What comes next

The document is a discussion paper rather than draft guidance. When the paper was released, CDRH Director Michelle Tarver, MD, PhD, described the goal as "a regulatory approach that keeps pace with the rapid innovation of digital health technologies," according to the FDA's announcement.