On October 15, JAMA published the first large look at how clinicians use an AI clinical information platform. Patel, Finlayson, Jena and colleagues, three of them from the company OpenEvidence, studied 952,008 verified US clinicians, 556,200 of them physicians, who sent about 271 million questions between January 2024 and June 2026. Monthly active users rose from 355 to 703,265. Nurse practitioners and physician assistants asked the most questions per month. Relative to their share of the workforce, physicians within a decade of graduation used the platform most (a use ratio of 2.54), followed by residents and fellows (1.84). A third of physicians’ questions concerned treatment and management; fewer than one in seven concerned diagnosis. Depending on the specialty, between 35% and 88% of questions fell outside the clinician’s own field, the highest share in emergency medicine.
The paper describes use, not outcomes. It cannot say whether the answers were right, whether they changed care, or whether care improved. What it does show is who is forming the habit: the youngest clinicians, at the stage when clinical judgment is still being built.
What the brain does with an answer it expects to receive
A study from University College London, now in Scientific Reports, asks what happens inside the brain when we hand information to an outside store. Scarampi, Sachdeva, Tsai and Gilbert scanned 36 young adults while they viewed sets of pictures and then received one of three cues: remember them for a test, forget them because there will be no test, or offload them because the test will come with a reminder.
When participants were told to offload, their brains behaved almost as if they had been told to forget. The research team trained a computer model to tell apart, in the brain scans, the moments when participants had been told to remember from the moments when they had been told to forget. The model classified 76% of the offload moments as forgetting. When people were told to remember, the brain kept a readable trace of the pictures until the test. When they offloaded, or were told to forget, that trace faded until it could no longer be detected.
The authors read this as adaptive: a memory system with limited capacity sensibly stops duplicating what a reliable store already holds. The long-term effect is unknown, and habitual reliance, they write, could reduce unaided memory if the freed capacity is not put to other use.
They also note a detail that matters for medicine. The small residual activity after offload cues may come from people keeping some trace in order to check whether the reminder was accurate. Verification needs something of our own to verify against.
What it changes for those of us who study it
The two studies should be read together with care. Neither shows directly that clinicians are losing anything, but together they point to where the evidence is missing.
We evaluate clinical AI by the accuracy of the tool. We rarely measure the person using it. Three measurements would change that. First, unaided performance over time: how residents who consult first perform on the same problems without the tool, at the start of training and at its end. Second, verification: how often a clinician checks an answer against a source or against their own reasoning, and whether that habit survives as trust grows. Third, the edge of competence: questions outside one’s own specialty are where a consult helps most, and where the clinician is least able to recognize a wrong answer.
None of this argues for keeping the tool from trainees. A reliable store may free the mind for other work, but we have not yet asked the critical question: what is that new work, what does it amplify in us, and are we measuring it?
So, for those of us who design these studies: what would we have to measure to know whether a generation that consults the machine first is learning less, or learning differently?
Sources. Patel VR, Finlayson SG, Liu-Hass M, et al. Clinician Use of an Artificial Intelligence-Enabled Clinical Information Platform. JAMA, published online October 15, 2026 · Scarampi C, Sachdeva C, Tsai P-C, Gilbert SJ. Cognitive offloading weakens the neural representation of to-be-remembered information. Scientific Reports, 2026