The Half Billion Dollar Question Nobody Has Answered: Does Clinical AI Make You a Better Doctor?

By the numbers, clinical AI has already won. OpenEvidence reports more than 750,000 verified US physicians, daily use by over 40% of the country's doctors, and upwards of 20 million consultations a month. It closed a $250 million Series D in January at a $12 billion valuation and now sits inside Epic at major health systems. Google's AMIE, built on Gemini, has published Nature papers showing it matching or exceeding primary care physicians in simulated multi-visit OSCEs. The tools are in our pockets, our workrooms, and increasingly our EHRs.

Here is what has not happened: a single rigorous study demonstrating that any of these platforms makes a physician better at medicine.

That is not rhetoric. A systematic review published this year searched MEDLINE, Scopus, Web of Science, and Google Scholar from inception through January 2026 for every evaluation of OpenEvidence — the most widely adopted clinical AI tool in American medicine. It found eleven studies. Eleven, against tens of millions of monthly consultations. And the eleven are heterogeneous in their questions, reference standards, and metrics; most are small, single-specialty, or built on simulated scenarios. There are no randomized trials of clinical outcomes, no formal validation against expert consensus. The literature confirms the platform is being studied. It does not show it has been validated — and it says nothing at all about whether the physicians using it are gaining anything durable.

The distinction that matters is between performance and capability. The best trials in this space measure what a physician does with the tool in hand. Some of those results are genuinely encouraging: a 2026 randomized trial found that physicians given LLM access after twenty hours of AI-literacy training scored dramatically higher on diagnostic vignettes than colleagues using conventional resources. But that is a measurement of the human plus machine dyad at the moment of use. It tells us nothing about what happens to the human's own knowledge base and clinical reasoning over months of daily reliance. No study of OpenEvidence, Gemini, ChatGPT, or any peer platform has measured that. It is the question every one of us should be asking, and it is the one question the evidence base has skipped.

Where longitudinal, real-world data does exist, it should give us pause. A multicenter study across more than 23,000 colonoscopies found that after several months of routine AI-assisted detection, endoscopists' unassisted adenoma detection rate fell from 28.4% to 22.4% — a roughly 20% relative decline in a skill that directly determines who develops colon cancer. It is the first real world documentation of clinical deskilling from AI, and a 2026 scoping review found the same signal echoed across pathology and decision support, alongside consistent evidence of automation bias: in one experiment, over 30% of clinicians reversed their own correct diagnoses when an AI confidently suggested a wrong one. Even physicians who had completed formal AI training accepted deliberately planted diagnostic errors at troubling rates. And in a randomized simulation trial of trauma decision making, teams with LLM access were no more accurate, slower on complex cases, and discussed less than teams reasoning on their own.

So we are left with a strange asymmetry. The adoption curve is vertical. The validation curve is flat. And the only evidence we have about long-term effects on the clinician points toward erosion, not growth.

None of this is an argument against these tools. Point-of-care evidence retrieval is genuinely useful, and I use it. It is an argument about direction of dependence. The current generation of clinical AI is built on a quiet assumption: that the model reasons and the physician executes. Every workflow that positions us as verifiers of machine output — rather than as reasoners whose judgment the machine amplifies — trades a little of our independent capability for convenience, and the colonoscopy data suggests the trade is real. Twenty years of pattern recognition, judgment under uncertainty, and hard-won intraoperative instinct do not live in a retrieval index. They live in us — and like any skill, they atrophy without use.

The burden of proof should sit where it sits for every other intervention in medicine. Before we accept that daily AI consultation is making us better clinicians, someone has to show it — with the same rigor we would demand of a drug. Until then, the honest summary of the literature is this: these tools are verified to be popular. They are not verified to make you better. Protect the reasoning. It is the one asset no platform can give back to you once it is gone.

References: OpenEvidence Clinical Question-Answering Platform: Systematic Review of Early Evaluations. 2026. Budzyń K, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025. Qazi IA, et al. Large language model diagnostic assistance for physicians in a lower-middle-income country: a randomized controlled trial. Nat Health. 2026;1:198–205. Goh E, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open. 2024. Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians. ESMO Real World Data and Digital Oncology. 2026. Gustafsson J, et al. Integrating large language models into trauma education for medical students: randomized controlled pilot trial. JMIR Med Educ. 2026;12:e79134.