A discussion paper isn't guidance

On August 18, 2026, CDRH’s Digital Health Center of Excellence published Considerations for the Regulation of Generative AI-Enabled Medical Devices — twenty-six numbered questions on risk assessment, premarket evaluation, and postmarket monitoring for devices built on generative AI, filed under docket FDA-2026-N-7874. Comment closed October 19, a week before this post. The paper itself says, in substance, what every law-firm summary of it repeats: it is not a draft or final guidance, and it proposes no binding policy change. That line is doing more work than a disclaimer usually does, because it places the document outside a specific body of federal regulation — not beneath it.

The apparatus a guidance carries, and this didn’t

21 CFR 10.115, FDA’s own good guidance practices rule, governs guidance documents specifically: a defined term, with a defined process attached. A draft guidance is announced by a Federal Register notice of availability, and the public comments through a docket FDA opened for that purpose. FDA staff may depart from a finalized guidance only with justification and supervisory concurrence — meaning even a non-binding guidance still binds the agency to a process, if not to an outcome. None of that apparatus attaches to a discussion paper, because a discussion paper is not the thing 10.115 defines. FDA announced this one by press release and a comment docket, the same two tools it would use to solicit feedback on anything at all.

A draft guidance is a promise with an open question in it: will it change before it finalizes. A discussion paper has no corresponding promise to keep.

The same week, the same center, a different track

The contrast is close at hand. CDRH’s FY2027 guidance agenda, published October 1, lists a draft guidance on generative-AI-enabled conversational devices aimed at mental disorders as one of this fiscal year’s commitments. When that draft lands, it will carry everything the August discussion paper didn’t: a Federal Register notice, a comment-and-disposition cycle under 10.115, and an A-list line item FDA already told the public it intends to work. The discussion paper asked a broader set of questions first, about generative AI devices generally, and closed its own comment window with none of that structure attached. Whether anything in the 26 questions informs the mental-health draft, a different guidance FDA hasn’t named yet, or nothing publicly traceable at all is FDA’s call to make, on no clock any rule sets.

Why it never became an event to watch

FDA Radar’s Federal Register source reads structured documents from federalregister.gov and sorts them by title: a document titled with both “guidance” and “availability” files as FDA Guidance; one with “classification of” files as a Classification Change; everything else defaults to a plain notice. A discussion paper announced by press release never reaches that feed at all — there is no Federal Register document for our source to fetch, let alone categorize. That is not a gap in our ingestion. It is the correct result of reading a format that, by FDA’s own description of the document, was never built to carry this one.

Worth reading anyway

Non-binding is not the same as uninformative. The paper previews a two-axis way of sorting risk that is worth running your own generative-AI-enabled function through now, obligation or not: one axis for how independently the function acts, from an informational output toward action-directing content, supervised action, and full autonomy; the other for the consequence of relying on an output that turns out to be wrong. A framework FDA is currently thinking out loud about is usually a better preview of tomorrow’s binding one than silence would be. Reading it, and deciding what it means for a specific device program, is a regulatory judgment call for your own team to make — this is regulatory intelligence and method, never regulatory advice, and least of all here, where even FDA has not yet decided what kind of document comes next.