FDA Wants Clinician Feedback on How to Regulate Generative AI Devices. Comments Close Oct. 19.
FDA's Aug. 18 discussion paper floats a two-axis risk framework, competency-based premarket testing and clinician review after launch for generative AI devices. It is not guidance. Docket FDA-2026-N-7874 closes Oct. 19, and clinicians are among those FDA invited to respond.
On Aug. 18, the Food and Drug Administration issued a discussion paper on how it might regulate medical devices that use generative artificial intelligence, and asked for feedback. The comment docket, FDA-2026-N-7874 on Regulations.gov, closes Oct. 19, one week from today.
FDA named clinicians among the groups it wants to hear from, alongside device manufacturers, consumers and researchers. A physician who works with an AI triage tool or a chatbot that answers patients can describe failure modes a manufacturer may not see.
What FDA published and what it did not
The paper, from the Center for Devices and Radiological Health, says it is for discussion only and does not represent draft or final guidance or propose policy changes. It poses 26 numbered questions.
It starts with a two-axis risk framework. One axis is how independently a function directs or takes action, from non-directive information to fully autonomous action. The other is the severity of harm if the output is wrong. The first question asks whether other dimensions belong in the framework, such as reversibility of a resulting action, downstream safeguards, time pressure and traceability to source material.
Testing a device the way clinicians are tested
For premarket review, FDA describes a competency-based approach "inspired, at a high level, by how human clinicians are evaluated and credentialed." It has two parts, non-clinical benchmarking and clinical confirmation, and would evaluate the final user-facing device as deployed, not the foundation model standing alone. The scope of testing would be proportionate to a device's risk.
After market, the paper lists three example approaches: periodic re-benchmarking, periodic sample-based review of real-world inputs and outputs by qualified, independent clinician adjudicators, and monitoring for performance drift. It also asks whether FDA could accept more premarket uncertainty if postmarket monitoring is strong, and what roles clinicians and health systems should play without diluting manufacturer accountability.
Nothing in this paper binds a manufacturer or a clinic. It is a chance to tell FDA what these tools do in the exam room before the agency decides what to require.
The paper notes that agentic systems used for documentation, care coordination and patient outreach may not be functions FDA focuses its device oversight on, so it does not settle whether a given scribe or inbox tool is a regulated device. Our earlier piece on governing ambient documentation tools covers what a practice controls either way.
What the docket shows so far
Regulations.gov listed 129 posted comments as of Oct. 12. A census by the regulatory consultancy Innolitics, coded with language models, counted 107 comments from 96 submitters as of Sept. 21. Recurring requests were to weigh reversibility of harm, evaluate the deployed system rather than the model alone, keep a human before irreversible actions and keep accountability with the manufacturer.
What to tell FDA by Monday
Useful comments are specific. Describe errors you catch in review, such as omitted history or a triage disposition set too low, how much review time is realistic and who would do it. Report whether a tool's behavior changed after an update you were not told about; question 24 asks how manufacturers should handle changes made by a third-party foundation model developer.
Tie each point to a numbered question, keep it short and leave out patient details, since the docket is public.
Reported, fact-checked and edited by The Script Pad's editorial staff.
This article is for professional education and does not replace clinical judgment. Treatment decisions should be based on the individual patient and current guidelines.
