In mid-2026, MedTech Dive downloaded the FDA's public list of artificial intelligence and machine learning-enabled medical devices and counted more than 1,350 entries — roughly double the count from just three years earlier. The FDA itself has maintained this list since 2021 and updates it periodically (most recently in March 2026); it is the closest thing to an official scoreboard for how much AI has actually entered clinical practice under FDA authorization. The number is real, it is growing fast, and it is also widely misread.
The first thing the list reveals is the pathway. The overwhelming majority of these devices — well over 95 percent — were cleared through the 510(k) route, which requires a manufacturer to show "substantial equivalence" to a device already on the market rather than proving safety and effectiveness from scratch. De Novo authorizations, used for genuinely novel device types with no predicate, and full Premarket Approval (PMA), reserved for the highest-risk Class III devices, together account for only a small fraction. That matters: 510(k) clearance is a comparative bar, not an efficacy bar in the way people assume regulatory clearance works.
The second thing the list reveals is specialty concentration. Peer-reviewed analysis of the FDA's 2024 authorizations found radiology accounted for roughly three-quarters of new clearances that year, with cardiovascular and neurology devices a distant second and third. Entire clinical domains — primary care triage, mental health, most of surgery — are barely represented. "AI is transforming medicine" is true mostly of image-heavy specialties where pixel-level pattern matching maps cleanly onto existing regulatory categories for computer-aided detection and diagnosis software.
The third and most misunderstood point is what "AI-enabled" means at the point of clearance. Under the FDA's traditional framework, a cleared algorithm is locked: its weights and outputs are fixed at the moment of authorization, and any meaningful change to the model requires a new submission. The self-improving, continuously-learning system that adapts to new data in the field — the thing most people picture when they hear "medical AI" — has historically not been how these devices operate once they reach a hospital. The FDA built its clearance process around software that behaves identically on day one and day one thousand.
That is what the Predetermined Change Control Plan (PCCP) framework, finalized by FDA in a December 2024 guidance document, is designed to change. A PCCP is submitted alongside the original 510(k), De Novo, or PMA application and specifies, in advance, exactly which future modifications a manufacturer intends to make — new training data, expanded patient populations, updated performance thresholds — along with the validation methods FDA pre-authorizes for each one. If a manufacturer sticks to the plan and its impact assessment, later versions of the algorithm can ship without a fresh marketing submission. It is less a leash removed than a leash pre-measured: the FDA is not approving open-ended learning, it is approving a bounded, verifiable range of future states.
Two limitations deserve equal billing with the good news. First, adoption of PCCPs is still early; most of the 1,350-plus devices on the list were cleared under the old locked-algorithm model and have no PCCP at all, so the framework's real-world track record is thin. Second, and more structurally, FDA clearance is a point-in-time judgment, and the agency has no comprehensive, mandatory system for monitoring how a cleared algorithm actually performs once it meets messier real-world data — different scanners, different patient mixes, different sites — than the validation set it was cleared on. Post-market performance drift for AI devices is discussed extensively in FDA guidance and academic literature, but it is not tracked with anything close to the rigor of the premarket review itself.
The count keeps climbing: by the close of 2025 the FDA's list had reached roughly 1,451 authorized AI/ML-enabled devices, with radiology still accounting for about three-quarters of the total. PCCP adoption, however, remains hard to quantify — the FDA has not published a running tally of how many cleared devices actually carry an approved change control plan, and industry coverage through mid-2026 still frames uptake as early rather than mainstream. Separately, on January 6, 2026, the FDA issued revised final guidance on Clinical Decision Support Software and its General Wellness Policy for Low-Risk Devices, and a new Quality Management System Regulation aligning FDA requirements with ISO 13485:2016 took effect February 2, 2026 — both relevant to how AI device manufacturers document and maintain compliance going forward.
None of this makes the underlying trend less significant. A regulatory system built for fixed software is visibly retrofitting itself for adaptive systems, one guidance document and one PCCP at a time. The next few years will show whether that retrofit holds: whether PCCPs get used broadly rather than as a niche option for well-resourced manufacturers, whether FDA builds real post-market surveillance to match its premarket ambitions, and whether specialties beyond radiology start showing up on the list in meaningful numbers. The count of 1,350-plus devices is a milestone in volume. The PCCP framework is the first real attempt at a milestone in kind.