Doctors Picked Their Own AI. That Is the Breakthrough — and the Trap.

Last month, America's doctors typed more than 42 million questions into a search engine most patients have never heard of. That figure comes from OpenEvidence itself, and it deserves the discount any self-reported number gets. But the Miami startup has hospital systems like Mount Sinai and Memorial Sloan Kettering embedding its tool in their records, a $15 billion valuation, and nearly $1 billion raised in eighteen months — so the claim is at least pointing at something real, Ars Technica reports. And what it is pointing at is the adoption pattern every AI company has been chasing since the chatbot era began: no procurement cycle, no CIO mandate, no pilot program with a steering committee. Individual physicians, in one of the most gatekept professions on earth, chose a tool the way consumers choose apps — because it worked at the point of need.

That is the milestone, and I want to argue for it plainly. Enterprise AI has spent years trying to buy its way into serious work through top-down contracts, and the results have mostly been demos and dashboards. OpenEvidence went the other way. Paul Testa, chief health informatics officer at NYU Langone, told the FT's reporting that his system sought a formal tie-up only after discovering clinicians were already using the tool on their own. The institutions followed the users. In medicine, of all places — a field that made an incumbent software vendor like Wolters Kluwer wait until October 2025 to ship an AI version of its flagship product — a startup founded in 2022 lapped the market on word of mouth alone. If generative AI can earn pull demand from people whose errors kill, it can earn it anywhere.

The honest objection is that demand is not accuracy, and on accuracy the record is genuinely thin. A Mayo Clinic study found OpenEvidence gave accurate, evidence-based recommendations in all five sampled cases — five, and Mayo is an investor. A June paper in Nature found it underperformed general-purpose models like Google's Gemini and OpenAI's GPT on clinical benchmarks, Ars Technica reports. A neurosurgeon at NYU Langone notes there is little research on individual tools and scant guidance on deploying them clinically. Fair enough. But notice what this objection does and doesn't establish. It means the evidence base hasn't caught up to usage — not that the usage is fake. Doctors are not a credulous audience; they have reference tools, colleagues, and malpractice exposure. When a UCSF resident says the tool lets her add patient context and decide faster, and when hospitals race to integrate what their staff already trusts, that is a signal worth more than a benchmark table. The right conclusion is not that the adoption is hollow. It is that the measurement must now be made worthy of the adoption.

Which brings me to the trap, because the business model OpenEvidence chose is the one thing that could squander everything its adoption curve proves. The platform runs on advertising. It booked $100 million in ad commitments by the end of 2025 and is on track for $400 million to $500 million in ad revenue this year, according to a person close to the company, Ars Technica reports. Sit with that. A tool shaping diagnosis and treatment, paid for by advertisers whose names and influence the reporting does not disclose. Medicine spent a century building firewalls between commercial money and clinical judgment — around sales reps, around journal ads, around prescribing — because everyone in the field learned, at cost, what happens when the entity paying is not the entity deciding. OpenEvidence is now rebuilding that conflict inside the workflow itself, and expanding into ambient documentation, which puts it in the room during the patient encounter. The metric an ad business optimizes is engagement — more prompts, more sessions — and 42 million monthly prompts is an engagement number. The incentive gradient of the whole enterprise currently runs toward volume of use, not quality of outcome.

The company has also made a revealing choice about whose rules it will live under: it withdrew from the EU and UK earlier this year rather than operate under the bloc's AI Act, Ars Technica reports. A company confident in its evidence base walks into regulated markets. A company protecting an ad model walks out.

Meanwhile the endgame, as founder Daniel Nadler describes it, is literature-based drug discovery — clinical trials, novel hypotheses, an economic prize he says will dwarf whatever hospitals or advertisers could pay. OpenAI and Anthropic are circling the same space, which validates the category and raises the stakes. The physician-facing tool, on this telling, is the wedge: the trust and the data come first, the real business later. That makes the terms on which trust is being accumulated right now the most important part of the story, and the least examined.

None of this is an argument against what OpenEvidence has built. It is an argument that what it has built is too valuable to run on the wrong incentives. The fix is not exotic: independent evaluation at real scale, published accuracy data rather than five favorable cases, clear labeling and firewalls around advertising, disclosure of who is buying influence near clinical answers. Doctors have handed AI the most credible demand signal it has ever earned. The field should treat that trust as capital to be compounded — because the alternative, a generation of physicians trained to reach for a tool whose answers are quietly for sale, is a debt no breakthrough will repay.