A 510(k) clears because something enough like it already cleared. That is the whole mechanism. Here is the machine it built — twenty thousand clearances, eleven and a half years — and the reason the FDA just proposed examining generative models instead of comparing them.
The FDA’s Digital Health Center of Excellence put out a discussion paper on generative AI-enabled medical devices and opened docket FDA-2026-N-7874 for comment through October 19. Inside it is a premarket path built on competency assessment — non-clinical benchmarking followed by clinical confirmation — which the agency describes as “inspired, at a high level, by how human clinicians are evaluated and credentialed.”
The interesting question is not whether that analogy holds. It is why the agency needed an analogy at all. To see that, you have to look at what it is replacing.
Each circle is one FDA product code — a lineage of devices that clear off each other. Circle area is how many 510(k)s cleared under it since 2015. Horizontal position is the median days from submission to decision for that lineage. Rows are medical specialty, sorted by volume. Colour is filing momentum — red if more than 31% of the lineage’s clearances landed in the last three and a half years, which is what a flat filing rate would produce.
The x-axis is not a quality signal. Median review runs from 29 days to 290 across these lineages, and — usefully — that spread barely moves as you raise the sample-size floor: the interquartile range goes from 66 days at ≥13 clearances to 42 days at ≥120. So this is structural, not small-sample noise. But calendar days from receipt to decision includes every hold clock and every additional-information round, which means a slow lineage is usually one where FDA kept asking and sponsors took months to answer. It is not agency speed, and it is certainly not rigour. Any chart that ranks device categories by review time is selling you one of those two readings.
Concentration is the real structure. Push the floor to 120 and twenty-seven lineages remain — twenty-seven product codes out of 1,573 — holding 5,994 clearances, nearly 30% of everything cleared since 2015, across nine specialties. The machine is not 1,573 doors. It is a few dozen well-worn ones, and each is a chain in which only the most recent link is ever checked: a device can clear today against a reference product whose own predicate cleared in 2003, whose predicate predates the internet. Nobody re-litigates the ancestor.
What a predicate certifies is sameness, not performance. Substantial equivalence asks whether the new device has the same intended use and the same technological characteristics, or different characteristics that raise no new questions of safety and effectiveness. It does not ask whether either device is any good. That is a design choice, and it worked for eleven and a half years because the things being compared held still.
Below is the same axis with one circle added, in the only position the current machine can give it: outside every lineage. It has no product code, no reference device, and nothing to be substantially equivalent to. That is not a gap in the taxonomy — it is what the discussion paper is a response to.
The FDA is not writing the exam. Read the paper closely and the agency proposes grading against a panel of qualified clinicians whose consensus reflects the standard of care, or against a median clinician in practice — and then stops, because writing a clinical exam that separates a competent answer from a merely fluent one is not a regulatory skill. It is a clinical one. The 20% of this that determines the other 80% is who shows up to the docket before October 19. By the time it is draft guidance, “competency assessment” will mean whatever the loudest commenters said it meant.
It covers 510(k) only. De Novo and PMA devices are excluded by construction, which matters because that is where the genuinely novel things go — including most of what would count as high-risk AI. It counts filings, not products: one firm re-clearing the same trade name eight times is eight dots’ worth of circle area and one capability. Product codes are assigned by FDA reviewers and drift, so a lineage that looks dormant may have migrated to a newer code rather than died. And review time here is calendar days from receipt to decision, which includes every hold clock and every additional-information round — it is not FDA work time and should not be read as agency speed.