A large language model does not read a scan and drop it in a bin. It writes. Ask it the same question twice and it can hand back two different answers, and every so often it will state a plausible falsehood with the exact confidence it uses for a fact. That is not a defect the vendors are about to patch out of existence; it is how the technology works. So the document the FDA released on August 18 is a strange one. The agency has begun building the regulatory pathway for medical devices that run on precisely this kind of model, and it has never cleared a single one.
Start with the catalog we already have. A taxonomy of 1,016 authorized AI devices published in npj Digital Medicine sorted a decade of FDA clearances and found them dominated by analysis: 85.6 percent read an input, an image most of the time, and return a classification or a measurement. The machine sorts. It does not invent. A little over one hundred devices do “generate” something, but the generation is narrow, mostly sharper images, guidance for capturing a scan, and synthetic training data. When the authors went looking for the free-text, reasoning kind of system, they reported plainly that they “did not find evidence of large language models” anywhere in the list. Out of 1,016 clearances, the count of the hallucination-prone models the new paper is written about is zero.
The FDA’s Digital Health Center of Excellence issued its discussion paper on how it might regulate these systems and opened a comment window under docket FDA-2026-N-7874 that closes October 19, 2026. A discussion paper is a request for comment, not a rule and not a cleared product, and the agency is careful to say it is not proposing policy yet. What it does is sketch the shape of the eventual rules. It lays out a two-axis scheme for sorting devices by risk, then proposes a premarket check it calls competency assessment, drawing an explicit parallel to the way physicians are credentialed: benchmark the model against known cases first, confirm it in the clinic before it reaches patients.
The analogy is warm, and it is intuitive, and that is exactly why it needs a hard look. Credentialing carries a threat behind it. When a physician invents a finding, they lose their license and often their career, and every doctor practices knowing it. A generative model carries no such stake. It cannot be sanctioned and it cannot be deterred, and when it fabricates, the strongest lever the framework offers after deployment is a monitoring plan while the device keeps working. The paper does concede that these systems can perform differently over time, and it reaches for the harder questions too, accountability for the foundation models upstream and agentic tools that act with little human oversight. But a credentialing metaphor quietly promises an accountability the underlying technology has never had to earn.
We have watched a version of this before. The FDA already runs an on-ramp that lets conventional AI devices rewrite themselves after clearance, the predetermined change control plan, a pre-approved document a manufacturer uses to modify its model without filing anything new. A study in JAMA Health Forum examined how those plans are used and reported and found the public reporting thin, which means a clinician often has no way to see how a cleared device has changed since the day it was authorized. Now point that same self-update machinery at a model that composes fresh output on every query and shifts as it learns. Postmarket monitoring is only ever as good as what regulators and clinicians are actually allowed to see.
Then there is the framing. The FDA presented the effort as advancing the administration’s priority to harness AI and accelerate the delivery of innovative medical products to market, and CDRH director Michelle Tarver cast the approach as “a potential model for regulators around the world.” Faster delivery is the manufacturer’s interest before it is the patient’s, and a framework built to be copied tends to travel before any single device inside it has proven it works. This publication does not grade that on party lines. When any administration puts acceleration first for a technology whose defining trait is confident invention, the burden of proof belongs with the agency, not with the people asking it to slow down.
None of this makes the FDA wrong to start writing. The technology is arriving whether the rulebook is ready or not, and even the researchers pushing hardest for oversight call these frameworks an urgent priority and argue for scrutiny across the whole product lifecycle, bench to bedside and back. The field admits it is building the plane in flight. The open question is who the rules are built to protect. Comments close October 19. Watch two things after that: how far the final framework tilts toward speed over confirmation, and which company files the first large-language-model device application to test the lane it opens. That first product will tell us more than the discussion paper ever could.
Sources
- FDA – Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback
- FDA – Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices (Aug 18, 2026)
- AHA News – FDA seeks feedback on regulatory approaches for generative AI-enabled medical devices (docket FDA-2026-N-7874, comments due Oct 19, 2026)
- npj Digital Medicine – How AI is used in FDA-authorized medical devices: a taxonomy across 1,016 authorizations
- JAMA Health Forum – Use and Public Reporting of Predetermined Change Control Plans for AI-Enabled Medical Devices (2026)
- npj Digital Medicine – Innovating global regulatory frameworks for generative AI in medical devices is an urgent priority (2026)
- European Heart Journal – Digital Health – Total product lifecycle regulatory considerations for generative AI-enabled medical devices (2026)