The FDA publishes a spreadsheet of every AI-enabled medical device it has ever authorized. One row per device: a name, a company, a clearance date, a review pathway, a medical specialty. As of late March 2026 it holds 1,524 rows, and it is one of the strangest artifacts in the whole AI discourse, because it is the rare place where a decade of arguments about "AI regulation" collapses into something you can actually count.
So this piece counts. The commentary around this list makes three standard claims: the growth curve is a hockey stick, radiology dominates, and almost everything comes through a fast lane called 510(k). All three are checkable, and here is the honest headline: two of them check out cleanly, the third is understated, and the most interesting object in the whole dataset turns out not to be in the dataset at all.
One bookkeeping note first, because it sets the tone for everything else. If you go looking for the total, you will find two. The FDA's own list, as published in late March 2026, carries 1,524 entries. An independent tracker's cumulative count through December 2025 says 1,451. The 73-entry gap is not an error, exactly. The list is a living document; the agency adds entries retroactively as well as forward, so any total is a snapshot with a date on it, and any article that tells you "over a thousand AI devices" without saying which snapshot is quoting a number that was already moving when they wrote it down. Both figures here come from the sources named at the bottom, with their dates. Keep that habit in mind; it is the whole method of this essay in miniature.
The hockey stick is real. In 2015 the FDA authorized 6 AI/ML devices. In 2025 it authorized 295. That is roughly a 49x increase in annual volume across a decade, and it is the least surprising fact in the file. Everybody asserts the curve; the curve computes. Fine.
Radiology's share computes too, and that is the more interesting result. The commonly cited figure is "about three-quarters," and the cumulative numbers through 2025 say 1,104 of 1,451 devices: 76.1 percent. In 2025 alone, 255 radiology AI devices cleared via 510(k). So the folk number survives contact with the data, which is worth saying out loud in an essay premised on checking rather than repeating.
But sit with the pairing of those first two facts, because together they say something neither says alone. The field grew 49x, and its composition barely moved. A decade of explosive growth in "AI in medicine" did not diversify medicine's AI; it deepened one specialty's lead. Growth stories usually broaden as they scale — new entrants find new niches, the long tail thickens. This one didn't. If you build or buy in this space, the list is telling you that the regulatory, technical, and commercial machinery for getting an AI device cleared is overwhelmingly machinery for getting an imaging device cleared, and every other specialty is still traveling a much thinner path.
The pathway claim is understated. The standard line is that "most" AI devices arrive via 510(k), the premarket notification route. For machine-learning-enabled Class II devices authorized in 2024, the measured split was 94.6 percent via 510(k) and 5.4 percent via De Novo. That is not "most." 510(k) is nearly the only door.
Which means you cannot understand this list without understanding what that door is.
A 510(k) clearance does not ask "is this device safe and effective?" in the way outsiders assume. It asks: is this device substantially equivalent to a device already on the market (a "predicate") in intended use and technological characteristics, without raising new questions of safety or effectiveness? Demonstrate equivalence and you are cleared. The heavyweight route that runs its own clinical evidence from scratch is the exception; for AI devices, at one door in twenty, it is nearly a rounding error.
If you are a developer, you already know this structure intimately, because it is transitive trust. It is the dependency tree. Nobody audits the whole stack; you trust your package because it declares a dependency on a package you trusted last year, which was trusted because of what it depended on before that. The 510(k) system is a dependency graph of safety arguments, and a clearance is an edge in it: this device is vouched for by that one.
And that raises the question the essay brief called the sharpest one: what do the chains look like? An AI device cleared as equivalent to an earlier device, which was cleared as equivalent to an earlier one still — how far back do these ancestries run, and what sits at the root?
Here is the finding about the finding: you cannot compute predicate chains from the AI list, because the list does not contain the predicate field. Clearance date, pathway, company, specialty: yes. Ancestry: no. Predicate relationships live in the individual 510(k) summaries and the separate 510(k) database. Walking a chain means joining datasets and reading documents one hop at a time, because, as a 2021 PLOS One genealogy study put it plainly, "the full genealogy of ancestral relationships connecting devices cleared through the 510(k) framework is not reported (only the immediate predicates of a device are listed)."
Let that register. The regulator's own public record shows each device's parents, never its lineage. The genealogy is an inference — always, for everyone, including the agency's own readers. In dependency terms: every package declares its direct imports, and there is no lockfile. If you want the transitive closure, you crawl.
People have crawled. This is not a new idea, and pretending otherwise would be theft: the PLOS One team mapped predicate genealogies across cleared devices generally in 2021, and a study in The Lancet Digital Health built the predicate networks specifically for AI/ML devices cleared from 2019 to 2021. What those walks found is the part worth carrying to a general reader.
More than a third of cleared AI/ML devices trace back to a non-AI/ML device in the first generation. The equivalence argument crosses the category boundary in a single hop: a system built on a learned model, vouched for by a device that does not learn.
And the ancestors are old. In the Lancet study's 2019–2021 cohort, the most recent ancestor that itself carried an AI/ML component dated to 2008 in cardiovascular devices — and to 2001 in radiology and haematology. A radiology AI cleared in 2021 could sit in a chain whose last AI-bearing ancestor predates AlexNet by eleven years, the deep learning era in its entirety, and every architecture the 2021 device is actually made of. Whatever that 2001 device did, it was not doing what a 2021 convolutional network does. In radiology specifically, the study also found the AI/ML task changed repeatedly along predicate networks — a drift with its own name in the literature, "predicate creep," documented in a 2023 PLOS One case study of a robotic surgical system.
Bound this honestly, because the point is sharper when it is fair. "Substantially equivalent" is a statutory term of art. It does not mean "identical" and never claimed to; it means comparable intended use and characteristics raising no new safety questions, and each individual hop in these chains presumably met that bar. The question the genealogy work raises is not whether anyone broke the rules. It is what a chain of pairwise comparisons is capable of carrying across twenty years — whether "no new questions, twenty times in a row" can quietly accumulate into a very new answer. Any engineer who has watched a codebase drift one reasonable refactor at a time knows the shape of that question.
There is a second, quieter way the list undersells what is happening, and most commentary misses it entirely. Roughly 10 percent of 2025 clearances, per secondary summaries, included a Predetermined Change Control Plan — a PCCP. A PCCP lets the manufacturer specify, in advance and with the regulator's sign-off, how the model may be updated after clearance without filing a new submission.
Think about what that does to the spreadsheet. The list records a clearance date, one row, one moment. A device with a PCCP is licensed to stop being the exact device that was cleared, within pre-agreed bounds. The row is a photograph of something designed to keep moving after the shutter closed. This is, to be clear, arguably the correct regulatory answer to the awkward fact that ML systems retrain — the alternative is freezing models in 2019 amber or re-reviewing every weight update. But it changes what a clearance date means, and it means one in ten recent rows is best read the way you read a dependency pinned to a version range rather than a hash: accurate on the day, guaranteed to drift.
Everything so far is about the gate. The list's deepest silence (and the brief's own caution, which deserves to be kept) is that it records nothing about what happened after the gate. Clearance is not efficacy. A 1,524-row list of authorizations contains zero rows about performance in the field.
But an outcomes dataset does exist, and it puts a clock on that silence. A study of 950 AI-enabled devices authorized through November 2024 (reported in JAMA-affiliated coverage) found 60 devices associated with 182 recall events. Two findings stand out.
First, the skew: publicly traded companies made 53.2 percent of the devices and accounted for more than 90 percent of the recall events — a 5.9x association between public-company status and recalls. That runs exactly opposite the intuitive story, in which under-resourced startups cut corners and incumbent giants have quality systems. The coverage also reports that devices lacking clinical validation were recalled more often than those with trials or retrospective studies behind them — a direction I will state without printing the per-segment percentages that circulate, because the denominators in the secondary coverage are murky and the underlying JAMA paper is where those numbers should be read before anyone repeats them.
Second, the timing, which is the number that closes this essay's argument without overclaiming: roughly 43 percent of those recall events came within one year of authorization. That statistic says nothing about whether any individual clearance was wrong. What it says is that the gate and the outcome are separated by months, not decades — that when the answer-by-analogy misses something, the miss tends to surface fast, in a system the list never records. The spreadsheet's silence about outcomes is a short silence.
For a reader who builds, buys, or signs off on AI systems, medical or otherwise, the list teaches a portable discipline: when someone shows you an approval, ask what the approval is. Concretely, three questions.
Which door? "FDA-cleared" spans a 94.6-percent-of-the-time equivalence argument and a from-scratch evidence review, and the phrase does not tell you which you got. Every certification regime you will ever meet (SOC 2, ISO, app-store review) has the same split between demonstrated-directly and inherited-by-analogy. Find out which one you are holding.
What predicate? An equivalence argument is an inheritance, and inheritances have ancestors. For a medical device, the immediate predicate is public record even though the full chain is not. For anything else (a model card citing a benchmark, a vendor citing a framework) the same question applies: what, exactly, is this thing claiming to be like, and how old is that thing?
What happens after the gate? Is the artifact frozen or licensed to change (the PCCP question), and is anyone measuring outcomes on the other side (the recalls question)? A green light with no post-gate telemetry is a very different object from one with a feedback loop, and the two look identical on a certificate.
And if you are the data-minded sort: the predicate networks have been walked for 2019 through 2021, and, in everything surveyed for this piece, nobody has published the walk for the 2024 or 2025 cohorts, the two biggest years in the list's history. The documents are public. The chains are sitting there, one hop per PDF, waiting for someone with a scraper and a free weekend. That is not a metaphor for anything. It is just true, and it is the most concrete invitation this spreadsheet extends.
The list itself will keep doing what it does: recording, one row at a time, a decade of the question "is it safe?" being answered by pointing at a previous answer. It is a genuinely useful public artifact, more transparent than what most industries publish about their gates. Just read it as what it is — a record of doors opened, with the doors' reasons filed elsewhere and the rooms beyond them unrecorded.
Sourcing notes: every figure is from the named public sources; none was computed from the raw list, which was not downloaded and row-counted for this piece — a reader needing one authoritative total should count the published list on a named date. The 2001/2008 ancestors and the non-AI-predicate share are the Lancet study's, 2019–2021 cohort only. The PCCP ~10 percent figure is secondary-sourced, unverified against a primary. The per-segment validation-absence percentages circulating in recall coverage were deliberately omitted pending the primary JAMA paper; the recall counts, public-company skew, and ~43 percent first-year figure are as the cited trade coverage reports them.
The discipline this list teaches is provenance: not is it approved but what is the approval, what is it inherited from, and how old is that ancestor. The FDA's own record shows each device's immediate parent and never its lineage, so the chain has to be reconstructed by hand, one hop per document.
Chain of Consciousness keeps that record for agent decisions: what was read, in what order, to arrive at what was claimed. The ancestry is written down rather than inferred.
pip install chain-of-consciousness and npm install chain-of-consciousness