Clinicians told the FT they accept AI at finding and push back at deciding. The FDA has been drawing that exact line for eight years — in regulation numbers, where nobody reads it. Here is every AI-pathway 510(k) clearance, one dot each, sorted onto the five rungs of what the software is allowed to say.
Primary source:Sarah Neville, “Medical AI has a proof problem,” Financial Times, September 20, 2026 — clinicians accept AI in imaging and diagnostics while resisting it in documentation, treatment recommendation and patient communication. Survey figures from Wolters Kluwer, 355 US doctors and nurses, March 2026. Data: FDA 510(k) premarket-notification releasable database and device classification file, via MIMI Labs, vintage 2026-09-14 (decisions through 2026-09-04). Every clearance marked substantially equivalent under one of 19 product codes in the AI/CAD regulation families 892.2050, 892.2060, 892.2070, 892.2080, 892.2090, 892.2100 and 870.2380 — 382 clearances from 204 companies. Rung assignment follows the regulation each product code sits under. Limits at the bottom — read them.
In eight years of AI-pathway 510(k)s, 12 devices — 3.1% — were cleared under 21 CFR 892.2060, the rule for software that renders a diagnostic characterisation. 223 sit on the bottom rung, where the software may measure and segment but must not say anything about pathology at all. The line clinicians drew in the FT is the line the regulation already draws.
Every AI-pathway clearance, on its rung
Each dot is one clearance. Horizontal axis is the year FDA decided; the five bands are the rungs, in ascending order of what the software is permitted to assert. Hover any dot for the device, the company and its K number.
Clinical area — assigned by keyword from the device name (approximate)
Count
2004
In view382
Companies204
Top rung (diagnose)12
Share at top rung3.1%
80/20 lens
Flip the count to distinct companies. The “Notify” rung is the loudest number in clinical AI — 82 clearances — but they come from only 41 companies, and one filer holds 19 of them. Repeat filings against the same predicate are how a product line looks like a market. Before you cite a clearance count in a deck, divide it by the number of firms.
Where each specialty is allowed to stand
The same 382 clearances, cross-tabbed: clinical area against rung. The interesting thing is not that radiology dominates — everyone knows that. It is that each area sits on a different rung, and the pattern inverts between chest and breast.
What the rungs actually mean
These are not editorial categories. Each is a distinct regulation with distinct language about what the output may assert.
Rung
Regulation
The device may output
n
Measure
892.2050 / 892.2100
Segmentations, volumes, contours, a reconstructed image. No claim about pathology.
223
Notify
892.2080 / 870.2380
“Look at this case sooner.” A worklist flag. Explicitly not a diagnostic claim, and not for primary interpretation.
82
Detect
892.2070
A mark on the image: something is here. No characterisation of what.
42
Detect + characterise
892.2090
The mark, plus a score or likelihood attached to it.
23
Diagnose
892.2060
A diagnostic characterisation of the finding, in the clinician’s hands.
12
Yesterday’s clearance, on the ladder
Today’s Curbside question was whether an imaging vendor cleared for seventeen findings diagnoses them. The ladder answers it without an opinion: a2z Radiology AI’s earlier a2z-Unified-Triage (K252366) cleared under QAS — 892.2080, the Notify rung. Seventeen findings, zero diagnostic claims. Every one of those seventeen is permission to move a study up the worklist.
“Cleared for appendicitis” and “detects appendicitis” are one preposition apart and mean different things to the resident reading at 3 AM.Builder’s Briefing, September 22, 2026
The part that should worry you
The Wolters Kluwer survey found 74% of clinicians named deskilling as a top concern, and 27% reported any awareness of AI governance where they work. Every post-market monitoring plan names clinician review as the mitigation. The bottom two rungs — 305 of 382 clearances — are precisely the devices whose safety case rests on a human reading carefully afterwards. The rung that requires the most human attention is the rung that got the least scrutiny going in.
Where this dataset is thin — read before quoting it
This is a product-code proxy, not FDA’s AI device list. It captures 510(k) clearances under the AI/CAD regulation families. A device using machine learning cleared under some other code is invisible here, and a device under these codes need not use ML at all.
The most autonomous device in the field is missing entirely. LumineticsCore (formerly IDx-DR), which diagnoses diabetic retinopathy without a clinician reading the image, came through De Novo, not 510(k). So did several others. The top rung is undercounted, and the ladder cannot see the autonomous path at all.
Clearances are not products, and products are not installs. One filer holds 19 Notify clearances that are largely successive additions to a single product line. Nothing here measures whether a device was ever deployed, or used twice.
2026 is partial. The file runs to 2026-09-04. The ten new abdomen/pelvis findings a2z announced on September 21 are not in it. Any apparent 2026 decline is the file, not the market.
The regulatory verb is not the clinical act. A triage device is barred from making a diagnostic claim, and still changes which scan a radiologist opens first at 2 AM. That is a decision about attention. The regulation governs what the software may say, not what the workflow around it does — which is the gap the FT piece is actually about.
Small-n rungs are noisy. Drag the year slider forward and the top rung drops to single digits fast. Do not read a trend off twelve dots.