clinicians.build · interactive

The Accountability Object

The Federation of State Medical Boards just argued that a license isn’t a certificate of correctness — it’s a name to attach when something goes wrong. This is the federal database of those names.

Primary source: Chaudhry & Valentine Theard, STAT First Opinion, Aug 3 2026
Data: NPDB Public Use Data File via MIMI Labs · 193,023 paid malpractice claims, incidents 2004–2021

The FSMB’s president and board chair published their answer on August 3 to whether generative AI should be licensed to practice medicine. The answer is no, and the reason isn’t accuracy. Their argument: a license is a grant of authority tied to a human being who can be disciplined, sued, and named.

Nobody disciplines a checkpoint.

So look at where the naming actually happens. The National Practitioner Data Bank holds 1,911,185 disclosable reports covering 985,019 distinct practitioners. Every one is a person. Not one is a product.

Below is the paid-malpractice half of it, cut two ways at once: what was alleged (11 allegation groups) × what happened to the patient (10 injury-severity codes). Every one of the 110 cells is a dot.

≥ 1
Cells shown
Paid claims
Total paid
Mean / claim

One dot = one (allegation group × injury severity) cell. Radius scales with claim count. Payments are the midpoints of coded ranges, not exact dollars — the NPDB public file does not release actual amounts.

Drag “min claims per cell” to 200 and watch the top-right corner. The most expensive-looking cells in this chart are mostly the smallest ones. IV & blood products → brain damage shows a mean near $875,000 on 17 claims. It sits beside real signals built on thousands. That is what a benchmark leaderboard looks like before anyone filters it.

The four groups a decision-support tool sits inside

Clinical AI does not have its own category here. Whatever it touches, the claim files under one of these eleven headings, under a person’s name.

Allegation groupPaid claimsShareMean payment% death

Rows in red are the four groups a diagnostic or decision-support tool most plausibly touches. Together: 63.8% of paid claims and 60.5% of dollars, 2004–2021.

Two accountability machines, moving opposite ways

Paid malpractice claims are the money machine. Adverse licensure actions are the license machine. Over the same twenty years, one got quieter and the other got busier.

Paid claims by incident year (left axis) vs. adverse licensure actions by action year (right axis). Incident years 2020–2021 are shaded: NPDB reports arrive years after the incident, so the most recent incident years are structurally incomplete. Do not read the tail as a trend.

What this dataset can’t tell you

Why a builder should care

Two state legislatures — Idaho and Iowa — introduced bills this year to license “autonomous service providers” outside the medical board. Both failed. FSMB has now told boards to go re-read how their state defines the practice of medicine.

That definition is the API contract every clinical AI tool is written against, and 69 boards are about to start editing it independently. Meanwhile the accountability record stays what it has always been: a name, a date, an allegation group, and a number.

Your real users aren’t the people who need the answer. They’re the people who have to sign it.


Read the FSMB piece in STAT →