AI Device Authorizations Grew Twentyfold in Seven Years. 2026 Is the First Year That Didn't.

Ask where AI in medicine is heading and you'll get opinions. Ask the FDA's public record and you get counts. Every number below comes from openFDA, the FDA's own public API — and the exact product codes and query are published at the bottom, so you can reproduce all of it yourself.

Quick answer

Across the eighteen product codes we track for AI-enabled software, FDA authorizations grew from 6 in 2018 to 141 in 2025 — more than twentyfold. But 2026 has interrupted that run: 77 authorizations through July 8, against 81 in the same window of 2025. On a like-for-like basis the category is flat to slightly down for the first time since we have data.

The trend, year by year

Counting authorizations by FDA decision date (openFDA's decision_date) across eighteen product codes covering automated image processing, computer-assisted detection and diagnosis, triage and notification, and machine-learning-based notification software:

• 2018: 6
• 2019: 14
• 2020: 33
• 2021: 38
• 2022: 61
• 2023: 83
• 2024: 102
• 2025: 141
• 2026: 77 — partial year, decisions through July 8

Seven years, a more than twentyfold increase. The pace moderated along the way: the two-year multiple ran about 5.5x from 2018 to 2020, then settled near 1.7x from 2023 to 2025. Strong compounding, but decelerating compounding.

What 2026 actually shows

The 2026 figure is not comparable to the years above it, and comparing it to them is the easiest way to get this story wrong. It is a partial year, and the dataset behind it was last refreshed on July 13, 2026, with no decisions posted after July 8.

Compare like with like — January 1 to the same date, same eighteen codes — and the picture is this:

• 2024: 46
• 2025: 81
• 2026: 77

That is a 5% decline against a window that had grown 76% the year before. We are not going to call it a downturn on half a year of data: first-half share of annual volume has swung between 36% and 51% since 2019, and recent decisions are still posting. But the honest read is that the compounding visible from 2018 through 2025 has not yet shown up in 2026, and anyone claiming the curve is still bending upward is not reading the current record.

We will publish the answer when the record has it.

Where the volume actually is

The distribution across categories says as much as the growth. All 556 authorizations, by code:

• Automated radiological image processing (QIH): 309 — more than half of everything in these categories, ever.
• Radiological triage and notification (QAS): 86 — software that flags a suspected finding and routes the case.
• Prioritization software (QFM): 41 — reordering the worklist, not making the call.
• Medical image analyzer (MYN): 37 — the older, general image-analysis code.
• Computer-assisted detection/diagnosis for cancer lesions (QDQ): 36, plus computer-assisted diagnostic software for cancer (POK): 23.
• Fracture CADe/CADx (QBS): 10, and ML-based quantitative imaging (QVD): 1.
• The machine-learning notification family — ten codes FDA classifies under cardiovascular: 13 between them.

The pattern: the AI reaching the US market fastest is workflow-layer software — processing, flagging, prioritizing — sitting on top of established imaging, with a clean predicate story. Standalone diagnostic claims remain the exception.

Three caveats that belong next to the numbers

These are records, not products. The 556 counts authorization records; they represent 455 distinct device names. A successful product clears again for each new version or added indication, so counting records overstates the number of distinct products by roughly a fifth.

Not all 556 are 510(k) clearances. 551 were substantial-equivalence determinations; 5 were De Novo grants — a different pathway, for novel devices with no predicate, reviewed on their own evidence. That is why we say "authorizations" rather than "clearances" for the total.

This is our watchlist, not the FDA's definition of AI. The eighteen codes are our selection, and they undercount the field considerably. The FDA publishes its own AI-Enabled Medical Device List, which carried 1,524 authorized devices across 172 product codes when we checked it on July 26, 2026. Our eighteen codes are a focused window on the imaging and notification categories, not a census of medical AI.

Check our math

We don't want you to take our word for any of this, so here is the exact basis for every number above.

The eighteen product codes: MYN, QIH, QAS, QDQ, QBS, QFM, QVD, POK, QXO, QYE, QXX, SAT, SHP, SII, SIJ, SIK, SBQ, SFR.

The field: decision_date. The source: the openFDA 510(k) endpoint. A year count is one query — set the date range and read meta.results.total:

https://api.fda.gov/device/510k.json?search=product_code:(QIH+OR+QAS)+AND+decision_date:[2025-01-01+TO+2025-12-31]&limit=1

Data currency at publication: openFDA reported last_updated of 2026-07-13; the most recent decision in this set is 2026-07-08.

Individual clearance records are also searchable by hand at accessdata.fda.gov — useful when you want the underlying documents rather than the counts. If you want to know what a specific clearance permits a vendor to claim, that lives in the 510(k) summary, not the database row; we walk through how to read one separately.

We track these categories weekly and publish the movement, free and with the query attached. If you reproduce these counts and get something different, we want to hear about it — that is the point of showing the work.

Common questions