5%
of 500+ studies used real patient data
What the research found
The ARISE network's review of more than 500 clinical-AI studies found only about 5% used real patient records; nearly half relied on exam-style questions. The report names a 'deployment gap' between controlled performance and how systems hold up in real care.
Why it matters for providers
Most claims of 'physician-level' AI rest on tests that don't resemble how the tools are actually used on patients.
Original source
Stanford–Harvard — State of Clinical AI Report 2026 (ARISE)
A reputable institutional report or news outlet.