Generative AI medical devices: what FDA has put up for discussion

2 October 2026 Michael Engel RA & medtech

FDA outlines a competency-based assessment for medical devices that use generative AI. Not a requirement yet, but a clear signal of what manufacturers should prepare for.

Source: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (FDA), 18 August 2026

Note: I provide engineering services in regulatory affairs for medical devices.

Technical drawing: a calendar sheet with 19 October 2026 struck through and a 'CLOSED' stamp for the expired comment period, next to a discussion paper on generative AI with a radar chart of competency areas and three margin comments.

Updated 27 October 2026: The comment period closed on 19 October 2026, and the docket does not accept late comments. According to regulations.gov, 541 submissions were received, 129 of which are publicly posted (as of 27 October). FDA has not given a date for the next step. According to MedTech Dive, however, Rick Abramson, director of FDA’s Digital Health Center of Excellence, said on 20 October to expect draft guidance “in a matter of months”. FDA’s guidance list for fiscal year 2027 so far has a guidance on risk assessment for generative AI devices only on its “Under Construction” list.

What happened: On 18 August 2026, FDA published a discussion paper on medical devices that use generative AI. Comments were accepted until 19 October 2026 (docket FDA-2026-N-7874). FDA states explicitly that it is neither draft nor final guidance but a basis for discussion.

The problem: Generative AI, meaning models that produce text or images, is hard to test with the classic “known input, expected output” approach. Answers vary, the model can state wrong things convincingly, and the underlying foundation model often comes from a third party that keeps changing it.

What FDA is considering (according to an analysis by Cooley dated 28 September):

Risk on two axes: How autonomously does the device act, and how serious is an error if someone relies on its output? Patient-facing functions could be placed higher on the risk scale than those aimed at clinicians; FDA explicitly asks for feedback on this.
Assessment like a professional’s: First, non-clinical benchmarking across defined competency areas such as clinical proficiency, safety behaviour and generalisability. Then clinical confirmation, scaled to risk, ranging from retrospective data to randomised trials.
Postmarket monitoring: Repeated benchmarks, spot checks by clinicians and monitoring for performance degradation (“drift”).
Foundation model master files: Providers of foundation models could file confidential information with FDA for manufacturers to reference. This would not amount to FDA authorising the foundation model itself.

What you can do now as a manufacturer:

Place the device on both axes and document the reasoning.
Build a benchmark set with acceptance criteria that can be rerun later.
Review contracts with foundation model providers: can a model version be pinned, and will you be told about changes?
Plan postmarket monitoring from the start, not after clearance.
Read the public comments on the docket and check which objections also apply to your own device.

Context: The direction is visible, but nothing is binding yet. Anyone planning today should treat the proposals as a likely yardstick, not as a current requirement.

What does this mean for your team? Let’s talk it through. Request an intro call