clinicians.build · July 11, 2026

Cleared, Then Unwatched

The FDA has now authorized 1,451 AI-enabled medical devices. Every one cleared a point-in-time review. This week a former surgeon described a validated, FDA-cleared decision-support tool whose recommendations physicians had silently overridden 40% of the time for six months — and nobody noticed. Below, 130 real AI 510(k) clearances. See the gate. Then see what watches the device after it goes live.

Primary source: Healthcare AI's next challenge isn't adoption, it's reliability (SmartBrief, Jul 2026)
Data: FDA 510(k) premarket notifications, AI/ML-related clearances 2015–2025 · via MIMI Labs · headline count: FDA AI-Enabled Medical Device List

The setup
1,451
AI devices FDA-authorized (through Dec 2025)
76%
of them are radiology
0
carry a mandatory ongoing performance-monitoring rule

A 510(k) clearance answers one question: is this device substantially equivalent to something already on the market? It is a snapshot taken the day the paperwork clears — not a promise the model still works in your hospital eighteen months later, on your scanner, on your patients. Each dot below is one real AI clearance. Its year runs left–right; the days FDA spent reviewing it run bottom–top; color is the medical specialty.

The explorer

The gate everyone photographs — and the hallway no one does

Critical lens — hide specialties with fewer than 1 devices130 of 130 shown
↑ The non-radiology dots look like “AI is diversifying into cardiology, neuro, ortho.” Drag right and that story dissolves — those specialties are 1–3 devices each. The trend is a small-n artifact. One specialty (radiology) is the whole picture.
130
Clearances shown
—
Median FDA review (days)
—
Share radiology
The critical lens

What the review time doesn't measure

Median review here runs about four to five months — long enough to feel rigorous. But every day counted on that vertical axis is spent before a single patient is touched. It measures the strength of the pre-market file: the equivalence argument, the retrospective validation set, the labeling. It measures nothing about what happens after go-live — drift as scanners and populations change, alert fatigue, the 40% of recommendations a clinician quietly reverses.

The axis that isn't here. There is no “overrides logged” axis, because that data mostly doesn't exist. A 510(k) clearance requires no ongoing performance reporting, no override telemetry, no drift alarm. The FDA's own remedy — a Predetermined Change Control Plan — is voluntary and forward-looking. So the honest version of this chart has a second, invisible dimension: for all 1,451 devices, the post-market monitoring value is the same number. Zero. Click “Show post-market monitoring” above to see it.

Note on the data: “AI device” has no clean flag in the 510(k) file, so this is a keyword-identified sample (~130 of the 1,451) — even defining the list is fuzzy. It mirrors the real distribution: overwhelmingly radiology, concentrated post-2019.

Why it matters for builders
The newsletter's thesis — “cleared is a snapshot; deployed is a moving target” — is this chart's empty top-right. Clearance is a solved, crowded gate: 130 dots and counting. The unbuilt product is the hallway after it — the override log, the drift monitor, the audit trail that tells a CMIO whether the tool a physician stopped trusting six months ago is still quietly firing. Instrument the override from day one. That single number is the thing your competitor's clearance certificate can't give the committee.
Read the reliability story → Or read today's newsletter — “The 40% override nobody logged” →