Every dot is 100 paid malpractice claims. Watch them sort themselves into what was alleged, and then into what it cost the patient. Nothing here is a model. Everything here is a person.
On August 3 the Federation of State Medical Boards answered the question directly: generative AI should not be licensed to practice medicine. Not “not yet.” No. And the reason isn’t benchmark performance.
A license was never a statement that you’ll be right. It’s a name to attach when you’re wrong.
1 dot ≈ 100 paid claims. Grouping and severity codes come from the NPDB Public Use Data File; state-fund duplicate payments excluded; incident years 2004–2021.
There is no allegation group for “the algorithm suggested it.” A tool that helps with a differential, a treatment pathway, a dose, or a monitoring threshold produces claims that file under diagnosis, treatment, medication or monitoring — four of the eleven headings, and together 63.8% of paid claims and 60.5% of the dollars in this file.
Click What was alleged above: the red columns are those four. That’s the surface area. The report that gets filed still carries one name, and it isn’t the vendor’s.
FSMB has told state boards to go re-read how their own statute defines the practice of medicine. That sentence is the API contract every clinical AI tool is written against, and 69 boards are about to start editing it independently, without a comment period you’ll hear about.
The practical move is small and it’s available today: every clinical tool you ship should be able to answer one question in writing — who is accountable for this output, and what did they see before they accepted it? If you can’t write that paragraph, you’ve built something that borrows a stranger’s license without telling them.
Want the same data with brushable axes and a sample-size filter? Open the explorer →