clinicians.build · interactive · September 22, 2026 · built on Sarah Neville, Financial Times, Sep 20 2026

Watch the Ladder Collapse

Clinicians told the FT they accept AI at finding and refuse it at deciding. The FDA wrote that same line into six regulation numbers years ago. Press play and watch every AI-pathway 510(k) clearance sort itself onto the five rungs of what the software is legally allowed to say.

Primary source: Sarah Neville, “Medical AI has a proof problem,” Financial Times, September 20, 2026. Clinician figures from Wolters Kluwer, 355 US doctors and nurses, March 2026.
Data: FDA 510(k) releasable database + device classification file via MIMI Labs, vintage 2026-09-14 (decisions through 2026-09-04). All substantially equivalent clearances under 19 product codes in the AI/CAD regulation families 892.2050–892.2100 and 870.2380: 382 clearances, 204 companies. Limits at the bottom.
Read the FT investigation → Or read today’s newsletter →
382 clearances · one dot each · hover any dot
step through it yourself

The rungs are regulations, not opinions

Each rung is a distinct rule with distinct language about what the output may assert. A device does not choose its rung; its product code does, and the product code is negotiated with FDA before the clearance letter goes out.

RungRegulationThe device may outputn
0 · Measure892.2050 / .2100Segmentations, volumes, contours, a reconstructed image. No claim about pathology.223
1 · Notify892.2080 / 870.2380“Look at this case sooner.” A worklist flag — explicitly not a diagnostic claim, and not for primary interpretation.82
2 · Detect892.2070A mark on the image: something is here. No characterisation of what.42
3 · Detect + characterise892.2090The mark, plus a score or likelihood attached to it.23
4 · Diagnose892.2060A diagnostic characterisation of the finding, handed to the clinician.12
80/20 lens Before you argue about whether a tool is safe, find out which rung it cleared on — it is one lookup in FDA’s 510(k) database, and it tells you what the vendor is permitted to claim, which is usually narrower than the slide deck. Today’s example: a2z Radiology AI’s a2z-Unified-Triage (K252366) cleared under QAS — 892.2080, the Notify rung. Seventeen findings, zero diagnostic claims. Each of those seventeen is permission to move a study up a worklist, nothing more.
“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 broken control

305 of the 382 clearances — 79.8% — sit on the bottom two rungs, where the device measures or flags and a human does the interpreting. That is the safety argument for all of them: a clinician reviews the output.

In the March survey the FT drew on, 74% of clinicians named deskilling as a top concern with clinical AI, and only 27% reported any awareness of AI governance where they work. Three-quarters of the people the mitigation depends on are saying the tool erodes the skill the mitigation is made of. That is not an ethics finding. It is a control that names a dependency nobody is measuring.

Where this dataset is thin — read before quoting it

Want to interrogate the same 382 clearances yourself — by year, by specialty, by company, with the repeat-filer collapse switched on? Open the explorer →