The FDA is getting ready to write generative AI medical device rules, and the outline is already public. In August the agency put out a discussion paper on generative AI-enabled devices covering risk assessment, premarket evaluation and postmarket monitoring. The comment window closed on October 19 with more than 500 submissions, and Rick Abramson, who leads the FDA's Digital Health Center of Excellence, told the AdvaMed conference that draft guidance should follow in months, according to MedTech Dive.
Two questions decide how hard the generative AI medical device rules bite
The paper sketches a two-axis way to judge risk. One axis is how independently the tool acts, from giving information to taking actions. The other is how much harm a wrong output could do, as Fierce Healthcare summarized it. A tool that drafts a note a clinician reads and edits sits in a very different place from one that acts on its own output. That is a useful lens even for software that will never be a regulated device.
Testing that looks like training
For premarket evaluation, the FDA floats a "competency assessment" loosely modeled on how physicians are trained and evaluated: benchmarking outside the clinic, then clinical confirmation before the tool reaches patients. The paper also takes on foundation models and agentic systems, where the same underlying model can be pointed at many tasks. On the agency's priority list for fiscal 2027, a draft on generative AI devices for mental health conditions sits near the top, with a broader guidance on risk and evidence still in progress.
One test is not enough
The U.K. regulator is moving in the same direction. Jessica Paulsen of the MHRA said regulators may be over-relying on single point-in-time assessments that don't reflect how these products behave in real settings, since the models are context dependent and can change. The MHRA has moved to adopt 44 recommendations from an independent commission and plans guidance on managing changes to devices after they reach the market.
For teams building clinical software, the practical read is to decide on purpose where a tool sits on the autonomy axis, keep it on the information side until the evidence supports more, and treat evaluation as something that keeps running after launch. Where does your team draw the line between a tool that suggests and one that acts?