clinicians.build · interactive · august 22, 2026

“Operates in Parallel With
the Standard of Care”

Every FDA-cleared imaging and cardiac AI device is defined, in regulation, by what the clinician does afterward. Here are all 110 of them. Find the one that had to show the clinician's review helped.

Primary source: Emanuel, Baker-Butler, Khosla & Khosla, JAMA, Aug 17, 2026
Data: FDA 510(k) release file & device classification database via mimilabs (vintage 2026‑08‑17)

Four authors argued in JAMA on August 17 that autonomous AI will beat both physicians and physician-plus-AI hybrids by 2030 — and that the hybrid is the worst of the three, because human oversight of a good model can make the output worse. The AMA's CEO said no on the record two days later.

That argument is an empirical claim about a control we already mandate. So look at what we mandated. Every one of these 110 clearances sits under a regulation that describes a human reading the output. Not one of them required evidence that the reading works.

“Radiological computer aided triage and notification software is an image processing prescription device intended to aid in prioritization and triage… The device does not mark, highlight, or direct users' attention to a specific location in the original image. The device does not remove cases from a reading queue. The device operates in parallel with the standard of care, which remains the default option for all cases.21 CFR 892.2080 — product code QAS — 46 of the 110 clearances below

The ledger

Each dot is one 510(k) clearance. Horizontal: FDA decision date. Vertical: days from submission to decision. Colour: what the regulation lets the device do before the human sees it.

Filter the ledger
1+
Clearances shown
110
of 110
Distinct firms
63
applicants of record
Median review
134
days, FDA receipt→decision
Required reviewer evidence
0
clearances

The trend that isn't

Drag min clearances per firm to the right. The apparent story — “FDA review of imaging AI is getting faster” — is carried by a handful of repeat filers using the Special 510(k) pathway for incremental updates to a device already on the market. Set the filter to 5+ and the slope you were about to quote in a deck comes apart.

Five tiers, one assumption

FDA sorts these devices by how far into the read they are allowed to reach. The tiers are precise about the machine. They are silent about the human.

“…intended to identify, mark, highlight, or in any other manner direct the clinicians' attention to portions of a radiology image that may reveal abnormalities during interpretation… This device is not intended to replace the review by a qualified radiologist, and is not intended to be used for triage, or to recommend diagnosis.” 21 CFR 892.2070 — medical image analyzer (CADe) — the “mark the image” tier

“Not intended to replace the review” is a statement about who is liable. It is not a statement about whether the review catches anything. Curt Langlotz, asked in May what research exists on how clinicians and models should be configured together, said “almost none” — then cited work showing that when a model is deliberately corrupted, readers follow it and get worse.

We adopted human-in-the-loop as a safety control and never instrumented the control. Builder's Briefing — August 22, 2026

What this chart is not

Read the seams before you cite it.
110 devices is not the AI estate. Ambient scribes, sepsis models, prior-auth engines and every LLM in a clinical workflow are mostly not 510(k) devices at all. This is the regulated sliver — the part where someone at least wrote the oversight assumption down.
The cohort is keyword-derived. Twelve FDA product codes whose classification names contain AI / machine-learning / computer-aided / triage / notification language. AI cleared under a generic product code is invisible here.
Review days ≠ scrutiny. Special and Abbreviated 510(k)s are short by design. A 22-day clearance is a process fact, not a rigour verdict.
The file ends at 2025-12-17. The 2026-08-17 mimilabs vintage of the FDA release file carries no 2026 decisions yet, so the right edge of the chart is a data artifact, not a slowdown.
“0 required reviewer evidence” is a claim about the classification regulations (identification text and special controls), not about every individual submission's clinical data. Some sponsors ran reader studies. None were required to report whether the reader's override was right.

The number nobody sells

All of us can name a metric for model accuracy. None of us can name one for oversight quality. Every deployed product in this ledger logs its own output. Almost none log the override and the outcome that followed it — which is the only dataset that settles the argument both sides of the JAMA debate were guessing at.

⚠︎ AI-generated · not reviewed by a human · verify against the linked sources before relying on it.