AI medical devices: the claim a regulator can quote back at you

AI medical devices: two regulators, one question#

Foxes learn early that a hedgerow tells you who passed through it, and regulators read marketing the same way. In the space of three weeks, two of them have published their thinking on AI medical devices. On 18 August 2026 the FDA's devices arm, CDRH, released a discussion paper on generative AI-enabled medical devices, and the US Small Business Administration noted that comments are due by 19 October 2026. On 10 September the UK's National Commission into the Regulation of AI in Healthcare published its recommendations.

Dating matters here, so let us be exact. The FDA paper is six weeks old, not new. What is fresh is the analysis: Cooley published its alert on 28 September, and Pinsent Masons followed on 29 September with a view on the UK report. Those two pieces are the news hook; the paper and the report are the primary documents underneath them, and both are worth reading for yourself.

Status counts too. The FDA is explicit that its paper does not represent draft or final guidance and is not meant to communicate the agency's regulatory expectations. Cooley adds that CDRH discussion papers have historically foreshadowed the agency’s regulatory direction, which is the quiet reason vendors should read the thing rather than wait. It also poses 26 discussion questions, so there is plenty to answer. Treat it as a scent on the trail, not a rule in the statute book.

The UK Commission was set up by the MHRA in September 2025 and, according to the government, gathered evidence from more than 12,000 people over a year. The report's own Annex A lists Recommendations 1 to 44, so 44 is the figure used here. Both regulators are asking the same underlying thing: what does this AI product claim to be, and what does it actually do?

Why your marketing copy defines what the device is#

That ladder comes straight from the FDA paper, which walks through four versions of one lisinopril sentence and says directiveness may depend on the substance and context of an output, not only on words such as recommend or should. For a marketer, the lesson is blunt: the verbs on your landing page are evidence of what your AI medical devices are meant to do.

US law already says this. Under 21 CFR 801.4, intended use refers to the objective intent of the people responsible for the labelling, and the regulation says that intent may be shown by labelling claims, advertising matter, or oral or written statements by those people or their representatives. Advertising, sales decks and a founder's podcast answer are all in that sentence. Write them as carefully as the instructions for use.

The UK report reaches the same place from the other side. Its glossary defines intended purpose as the use stated on the labelling, the instructions for use and/or the promotional materials. Then it goes further. Recommendation 1 asks the MHRA to update the definition so that device design and functionality are appropriately considered in addition to a manufacturer’s claims and promotional materials. A conservative claim will not shelter a product that is built to do more.

That cuts both ways, and it is where the quarry gets interesting. Over-claiming draws attention, but so does under-claiming a product whose design says otherwise. The report notes it matters especially if that design is inconsistent with specific claims or promotional materials. The fox's rule: make the copy, the interface and the evidence tell one story.

Two of the Commission's transparency recommendations touch marketing directly. Recommendation 9 suggests guidance on model cards, careful user-centred interface design and dynamic labelling, and the government says the Commission's engagement found broad public support with clear conditions: strong safety standards, meaningful human oversight, and transparency about when AI is used. Plain, versioned statements about what a model does are about to become a regulatory habit, not just a brand choice.

Software as a medical device: where the FDA's two axes put your function#

The FDA's framework puts what a function does on the horizontal axis and the severity of relying on a wrong output on the vertical one, with risk rising towards the top right. CDRH calls it a possible organising heuristic, not a rule. It also says something that every vendor of AI medical devices should print out: FDA does not regulate GenAI as such. It regulates products that meet the device definition.

That is the old software as a medical device logic applied to a new thicket. Some software functions are not devices at all, some are low-risk and subject to enforcement discretion, and the rest need oversight. What the generative layer changes is the difficulty of describing the function: open-ended inputs, variable outputs and conversations that can drift. Cooley notes that CDRH also wants views on multi-turn chats that migrate from information to action-directing advice.

Audience is the next lever. The paper floats shifting patient-facing informational functions higher on the consequence axis than the same function for clinicians, because HCPs may be better equipped than patients to interpret an output in context. It also asks whether that shift is right, since patient empowerment has value. For a marketer the practical point is simple. The same feature of AI medical devices, described to a consumer, may carry a different regulatory weight than when it is described to a clinician, so write separate pages.

Cooley adds a sharp observation: the FDA's worries about measurement and signal-processing functions and about patient-facing output appear to come directly from concepts that are structural to its clinical decision support guidance. So the thinking is not wholly new: familiar risk logic, stretched over a technology that talks back.

FDA cleared vs FDA approved: what your badge is allowed to say#

Vendors love a regulatory badge, and the badge has a grammar. A 510(k) submission ends with an FDA order finding the device substantially equivalent, and the FDA's page says that order clears the device for commercial distribution. Premarket approval is a different, tougher route: PMA approval is based on a determination by FDA that the PMA contains sufficient valid scientific evidence to assure safety and effectiveness. FDA cleared vs FDA approved is no pedant's quibble; each word is a claim, and the wrong one reads as overreach.

For AI medical devices the practical question is which route the product took. We counted the entries on the FDA's AI-enabled medical device list: of 1,614, a full 1,553 carry a 510(k) number, 40 are De Novo and 21 are PMA. Roughly 96 in 100 are cleared, not approved. If your copy says approved and your file says cleared, a competitor's lawyer will notice before a regulator does.

The list itself carries a warning label. The FDA says it is not a comprehensive resource of AI-enabled medical devices, built mainly from AI terms in marketing authorisation summaries. It also says it will explore ways to tag devices that use foundation models, so a future version of the FDA AI medical devices list may show which products rest on a large language model. Plan on being findable that way.

A crowded list of AI medical devices means more rivals making similar claims, and the FDA is building ways to see them. Standing out on facts, such as the indication, the dataset and the date, will outlast standing out on adjectives.

How to substantiate an AI claim without writing a device claim#

Here is the folkfox view, offered as judgement rather than legal advice: treat every claim about generative AI in healthcare as if it will be read next to the product's design. Marketing teams that run a claims register, like a medical-legal review for pharma, will cope with whatever the 19 October feedback produces. Good claims for AI medical devices are specific, dated and supported; teams that rely on a footer disclaimer will not cope, because the FDA is weighing whether such a statement leaves an output just as directive.

The FTC adds a second filter. Its health products guidance tells marketers to look at the net impression conveyed by all elements of the ad, and says the FTC can obtain an order requiring that future marketing be truthful and substantiated. So your AI claim must clear two readers: the device regulator, who asks what it is meant to do, and the advertising regulator, who asks whether you can prove it.

Then comes the dull, decisive craft of evidence. The FDA paper's competency idea, benchmarking followed by clinical confirmation, is modelled less on product testing and more on how human clinicians are credentialed. If that survives the feedback round, a vendor's best marketing asset will be a dated benchmark with a named method. Start collecting it now, because a number without a method will sound like a vulnerability.

UK vendors have a different brief. Pinsent Masons reports that the Commission's ideas, including staged authorisations and real-world monitoring, could help Britain carve its own path on medical device regulation. The Commission calls its staged model L-plates for learner drivers. Marketing for an L-plate product should sound like one: supervised, evidenced, improving.

If you want a second pair of eyes on a claims register, our healthcare marketing team builds them with clinical and regulatory advisers, and our content marketing work turns evidence into pages that stay accurate. For the model-side questions, see AI consultancy, and for how answer engines will quote your claims, see SEO and GEO. Related reading includes our AI healthcare regulation trust test and medical device agency briefs pieces, and the FDA untitled letter lesson on how the agency reads an advert. Ready to talk it through? Start at contact us.

Frequently asked questions#