270 Million Queries Show Where Clinicians Are Already Using Medical AI

A JAMA study of OpenEvidence logs found that clinicians used the platform most often for treatment and medication decisions—the very tasks whose real-world performance remains largely unmeasured.

Haivian Discovery News Desk2026-10-21

A network of clinical questions converges on an evidence prism and branches into treatment, medication and guideline pathways. Original Haivian editorial artwork.

Direct answer

What does the JAMA OpenEvidence study show?

It shows rapid, sustained use of one AI-enabled clinical information platform by 952,008 verified US clinicians who submitted 271 million queries from January 2024 through June 2026. The study describes who used the tool and what they asked; it does not show that the answers were accurate, improved decisions or improved patient outcomes.

What happened

A large cross-sectional study published online by JAMA on 15 October 2026 offers an unusually detailed view of how clinicians use an AI-enabled evidence platform in practice. Vishal Patel, Samuel Finlayson and colleagues analysed OpenEvidence accounts verified through US National Provider Identifiers.

Across the period from 1 January 2024 to 30 June 2026, 952,008 clinicians submitted 271 million queries. Monthly active users rose from 355 to 703,265, while average monthly queries per active user rose from 18.0 to 39.5. Physicians accounted for 556,200 users, but nurse practitioners and physician assistants submitted more queries per active user-month.

The striking result concerns purpose. In a stratified sample of 38.8 million queries, 51.9% of physician queries concerned clinical decision-making. Treatment and management represented 33.1%, medication information 14.1%, and diagnosis 13.9%.

How AI was actually used

There are two distinct AI roles in the study. First, clinicians used OpenEvidence, which synthesises answers to free-text clinical questions with citations to medical journals and professional guidelines. Second, the researchers used a large language model to classify the purpose and medical topic of the 38.8 million sampled queries.

The authors report that the classifier agreed with clinician annotators at least as often as clinicians agreed with one another, but they still list LLM misclassification as a limitation. Topic labels are heuristic, and generalist questions can span several specialties.

Why it matters

Many prominent medical-AI evaluations emphasise diagnostic vignettes and exam-style recall. The observed use pattern points elsewhere: clinicians were asking about treatment, management, medications, evidence and questions outside their own specialties. Those are consequential tasks with changing guidelines, patient context and asymmetric costs of error.

The study therefore supplies a better target map for independent AI evaluation. A useful audit programme would sample the questions clinicians actually ask, preserve source dates, test citation fidelity, examine clinically meaningful omissions, and distinguish a plausible synthesis from advice that is safe under real constraints.

What remains unproven

This study measures use, not answer quality. The authors explicitly say performance on treatment selection, medication guidance and evidence synthesis remains largely unmeasured. The records do not establish whether a query occurred during a patient encounter, whether a clinician followed the response, or whether use changed patient outcomes.

The data come from one platform, cover only its main question interface and may not generalise to products with different content access. The study also cannot distinguish new information needs from substitution away from older resources or consultation.

Commercial relationships require attention. OpenEvidence supplied the deidentified data and participated in analysis. Finlayson and Ray were OpenEvidence employees, and Zack was its chief medical officer during the study. Those disclosures do not invalidate the analysis, but they belong in any independent interpretation.

Primary sources and version boundary

For evidence systems that need citation, output and risk claims tested against realistic use, see Haivian's AI research audit.

Bounded audit pathway

Need a claim examined before publication?

Haivian offers scoped research audits with explicit evidence, limitations and accountable human sign-off. Payment does not purchase a positive result or publication.

Explore research audits