AI at the Bedside: Study Tracks How Nearly a Million Clinicians Use OpenEvidence

Artificial intelligence has moved from the margins of medicine into the daily workflow of hundreds of thousands of clinicians, yet until now, remarkably little has been known about who actually uses these tools, how often they reach for them, and what they ask of them. A new study led by researchers at Mass General Brigham and Harvard Medical School offers one of the first large-scale, data-driven portraits of clinical AI adoption in the United States, drawing on an extraordinarily rich dataset: more than 270 million queries submitted by nearly one million clinicians on OpenEvidence, the most widely used AI-powered clinical information platform in the country. The findings, published in JAMA, paint a picture of a technology that has spread rapidly through the clinical workforce and is being used in ways that may surprise both its champions and its skeptics.

The research team, led by Vishal Patel, MD, MPH, and overseen by Anupam Jena, MD, PhD, both of Mass General Brigham and Harvard Medical School, set out to close a persistent knowledge gap. AI-enabled clinical tools now allow clinicians to type questions such as how best to explain cholesterol levels to a patient or what the outcomes of a specific clinical trial were, and receive synthesized, evidence-grounded responses within seconds. But while the platforms themselves have attracted enormous attention and investment, robust data on real-world usage patterns, which clinicians adopt them, at what career stages, and for which kinds of decisions, has remained scarce. Without such data, efforts to evaluate the safety, value, and appropriate integration of clinical AI have lacked an empirical foundation.

To build that foundation, the researchers analyzed platform activity between January 2024 and June 2026, a window during which the technology’s reach expanded dramatically. Active users, query volume, and user retention all increased sharply across the study period. By the end of the observation window, nearly 952,000 clinicians had submitted approximately 271 million queries through the platform. Those numbers make this analysis, by a wide margin, the most comprehensive examination to date of how a clinical AI tool is actually used in practice, moving the conversation beyond surveys and anecdote to a direct measurement of behavior at population scale.

One of the study’s central findings concerns who is using the technology. OpenEvidence was adopted broadly across the clinical workforce, but its most enthusiastic users were not, perhaps as some might expect, the most senior physicians. Instead, use was especially high among nurse practitioners, physician assistants, early-career physicians, residents, and fellows. This pattern suggests that AI information tools may be filling a particular need among clinicians who carry heavy patient loads, face broad and varied clinical questions, or are still building the deep experiential knowledge base that comes with decades of practice. It also raises intriguing questions about how the next generation of clinicians will be trained, and whether rapid, AI-mediated access to synthesized medical evidence will reshape the habits of an entire professional cohort.

Specialty mattered as well. The researchers found that use was higher among physicians in oncology and hematology, internal medicine, and family medicine, fields characterized by rapidly evolving evidence bases, complex treatment decisions, and an unrelenting stream of new trial results. Oncology in particular generates a torrent of new publications, making it a natural fit for a tool that can surface and synthesize trial data on demand. By contrast, use was lower among physicians in ophthalmology, radiology, orthopedics, and anesthesiology. The authors’ analysis of these differences offers a window into where clinicians currently perceive the greatest value in AI-assisted information retrieval, and where existing tools may not yet align with the information needs of particular disciplines.

Perhaps the most consequential finding relates to what clinicians are actually asking. Contrary to the popular narrative that AI in medicine is primarily a diagnostic aid, the study found that most queries related to treatment and management decisions rather than diagnosis. Clinicians, in other words, are not mainly asking the platform what a patient’s symptoms might signify. They are asking how to treat, how to manage, and how to guide care once a clinical picture is already established. This orientation toward treatment and management reframes the risk-benefit calculus of clinical AI: rather than replacing the diagnostic reasoning that sits at the heart of physician expertise, these tools are functioning more like an always-available clinical reference, capable of pulling together current evidence at the moment a decision is being made.

The second major pattern in query behavior is equally revealing. Clinicians frequently sought information beyond their own primary specialty. A family medicine clinician might consult the platform about an oncology question; an internist might look up material from a surgical literature. This cross-specialty information-seeking underscores a structural reality of modern medicine: generalist clinicians routinely confront questions that stretch beyond the boundaries of their training, and the evidence they need is scattered across thousands of journals. An AI platform that can bridge those boundaries in seconds represents a fundamentally different model of clinical information access than the traditional search-and-read workflow, and the study’s data suggest clinicians are embracing it precisely for that reason.

The scale and speed of adoption documented in the study carry significant implications for the future of evidence-based medicine. Clinical guidelines and primary literature have always lagged behind the questions clinicians face at the bedside. If large fractions of the workforce are now turning to AI synthesis tools as a first stop for clinical information, the way medical evidence flows into practice may be undergoing a structural shift, one with enormous potential benefits in speed and accessibility, but also with new dependencies and failure modes that the medical community has only begun to think through. The study does not measure whether the platform’s answers are correct, nor whether its use improves patient outcomes; those questions remain open and are likely to define the next phase of research in this area.

What the study does provide is an empirical baseline. By quantifying adoption rates, usage trajectories, professional and specialty patterns, and the purposes for which clinicians query a clinical AI system, the researchers have given policymakers, health systems, educators, and platform developers a shared factual starting point. Future efforts to evaluate clinical AI tools, to design appropriate oversight, to integrate them into electronic health records and clinical workflows, and to train clinicians in their responsible use can now be grounded in observed behavior rather than speculation. In a field where hype has often outrun evidence, that in itself is a meaningful contribution.

The study, titled

Subject of Research: Study characterizes clinicians’ use of the AI platform OpenEvidence

Article Title: Study characterizes clinicians’ use of the AI platform OpenEvidence

Article References: Study characterizes clinicians’ use of the AI platform OpenEvidence. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: characterizes, clinicians, platform, OpenEvidence, scientific research, peer-reviewed research