A Harvard-linked study analyzing query data from OpenEvidence, a clinical AI search tool used widely by physicians, found that doctors are turning to AI primarily to help inform treatment and management decisions rather than to arrive at diagnoses, offering a rare, data-driven look at how AI is actually being used inside clinical practice rather than how it’s marketed.
Why This Distinction Matters
AI diagnostic tools have drawn the most public attention and regulatory scrutiny, but this study suggests the more common real-world use case is narrower and arguably lower-risk: physicians already know or suspect a diagnosis and are using AI to quickly surface the latest treatment guidelines, drug interactions, or management protocols rather than asking the AI to determine what’s wrong with a patient in the first place.
How This Fits a Broader Pattern in Clinical AI Tools
The finding aligns with how other AI-enabled clinical products are being built: Doximity has integrated UpToDate, a widely used clinical reference, into its AI support tool specifically to help physicians retrieve current treatment information quickly, while Epic has rolled out Connectors that let its AI assistants retrieve information and complete tasks across other hospital software systems — both reflecting demand for AI as an information-retrieval accelerant rather than an autonomous decision-maker.
The Diagnostic AI Debate Happening in Parallel
Separate research has shown AI diagnostic performance can be genuinely strong in controlled settings — one Harvard and Stanford-linked study using emergency department cases from Beth Israel Deaconess Medical Center found an AI model correctly or near-correctly diagnosed patients 67% of the time at triage, compared with 50-55% for human specialists — but doctors’ real-world usage patterns suggest a cautious, hybrid approach is winning out over full diagnostic delegation.
Why Physicians May Be Making the Right Call
A JAMA Network Open study published earlier this year found large language models tend to converge prematurely on a single diagnosis when faced with clinical uncertainty and competing possibilities, concluding these models aren’t yet ready for autonomous clinical decision-making — a limitation that treatment-support use cases largely sidestep, since the physician has already done the harder diagnostic reasoning before consulting the AI.
What This Means for AI Vendors
Companies building clinical AI tools may find more near-term commercial success and physician trust by focusing on treatment-support and information-retrieval use cases rather than competing to replace diagnostic judgment, a positioning that several vendors, including Doximity and Epic, already appear to be adopting.
What’s Next
Watch for expanded studies tracking how physician AI usage patterns evolve as diagnostic AI tools improve, and watch whether regulatory frameworks being drafted by the FDA distinguish between these two use cases, treatment support versus diagnosis, with different levels of oversight.
Photo: Tima Miroshnichenko / PEXELS via Pexels
