Study examines 271 million clinical questions submitted to AI
A JAMA study tracks OpenEvidence use by more than 952,000 US health professionals. Physicians’ questions focus largely on treatment decisions, but the research does not measure patient outcomes.
Summary
AI is already part of how hundreds of thousands of US health professionals seek clinical information. A study published in JAMA on October 15 examined 271 million questions submitted to OpenEvidence by 952,008 clinicians between January 2024 and June 2026.
The platform answers clinical questions with references to scientific papers and professional guidelines. Researchers at Harvard Medical School and Brigham and Women’s Hospital investigated who uses it and why, rather than simply testing whether a model can pass medical examinations.
In practice
Among physicians, 51.9% of classified questions concerned clinical decisions. Treatment and management accounted for 33.1%; diagnosis, 13.9%. Other requests involved clinical knowledge, evidence, administrative work and education. This distribution suggests that evaluating medical AI only on diagnostic performance misses a substantial part of actual use.
Questions frequently crossed specialty boundaries. Surgeons directed 70% of their questions to nonsurgical subjects. Early-career physicians used the platform more intensively, while monthly active users increased from 355 to 703,265 during the study period.
To categorize the content, the researchers analyzed a sample of 38.8 million questions with a language model and compared a subset of classifications with clinician assessments. This approach reveals patterns at scale, although classification itself has a margin of error.
What we still don't know
The study covers one platform in one country. It does not establish that the answers improved treatment, reduced errors or changed decisions. A question about a patient does not reveal whether the answer was followed. Harvard Medical School identifies examining effects on care as the next step.
OpenEvidence supplied de-identified data, and three coauthors have or had employment ties to the company. Disclosed funding includes support from the US National Institutes of Health. These details help readers interpret the research and its limitations.
Why it matters
- Clinical adoption can be studied through usage records alongside surveys and laboratory tests.
- Treatment, medication and questions outside a clinician’s specialty need attention in quality and safety evaluations.
- Widespread use makes research on patient outcomes urgent; it does not replace that research.