The volume of medical research is growing so fast that no single physician, or even a pharma R&D team, can realistically keep up with the frontier. Research still takes an average of 17 years to reach clinical practice. A new partnership between Wiley and OpenEvidence aims to narrow that gap by licensing more than 400 peer-reviewed journals, including the Cochrane Database of Systematic Reviews, into the medical AI platform used by over 40% of U.S. physicians.
OpenEvidence, founded in 2022 by former Kensho CEO Daniel Nadler, has raised roughly $750 million and reached a $12 billion valuation as of January 2026. The platform is free for verified physicians and supported by pharmaceutical advertising, and claims more than 18 million clinical consultations per month. That scale is part of what makes the Wiley deal significant.
While much has been made of AI’s potential in medicine, generalist AI systems have inconsistent accuracy: For instance, LLMs routinely ace medical board-style exams, but performance drops nearly 40 percentage points when questions are converted to free-response. LLMs, too, consistently underperform expert clinicians on the kind of open-ended reasoning physicians face daily. While AI systems can impress when they are accurate given their efficiency at data processing, “powerful AI tools without powerful content are hamstrung,” said Josh Jarrett, senior vice president and general manager of AI growth at Wiley. While AI gets a lot of attention on its own, it’s only half the story. “It’s actually much closer to 50–50 between AI technology and the trusted content and sources it draws from,” Jarrett said.
Jack of all trades, master of some
Wiley’s own ExplanAItions Survey found that roughly 84% of researchers were using AI in some part of their work in 2025. But increasing adoption is not leading to higher trust in generalist models. Jarrett said researchers are telling Wiley that the current generation of off-the-shelf “deep research” tools are broad rather than genuinely deep. “It searches lots of websites, but they want to go deep into proprietary datasets,” he said. “An engineer working on a bridge in Germany doesn’t care about what the web says, they care about the specific bridge guidelines for Germany.
That observation cuts to the core of why Wiley sees the OpenEvidence partnership as a template rather than a one-off. Under the agreement, Wiley will license more than 400 journals and books to OpenEvidence, including the Cochrane Database of Systematic Reviews, widely considered the gold standard for evidence syntheses used to inform clinical guidelines, along with Cochrane Clinical Answers, a curated collection of actionable clinical insights drawn from Cochrane reviews. The content spans specialties from cardiology and oncology to neurology and psychiatry, and will be cited transparently inside OpenEvidence’s evidence layer. Wiley says the content feed will update within a day of publication to its online library.
A retrieval system
Jarrett said the deal aims to improve what physicians can retrieve at the moment they need it. He described OpenEvidence as a retrieval-based system, essentially a RAG (retrieval-augmented generation) architecture, that uses AI for search as much or more than for summarization. “They summarize the results and say, ‘We found the right evidence from this source,’ but then they show the quote, the chart, the snippet from that original source,” Jarrett said. “There’s a live link. You can click back to see the full text if you have access. They’re not trying to just write a summary out of a generative AI engine.”
That distinction matters in the context of how Wiley thinks about trust in the AI economy. “We think trust becomes the scarcest resource in the AI economy,” Jarrett said. “You see three citations, a snippet you can read yourself, you can click through, and it’s from brands that have independent trust within that problem space.”
It takes a village
That setup also fits Wiley’s broader AI strategy. Instead of building a closed assistant of its own, the publisher is placing its content inside multiple AI tools that researchers and clinicians already use, provided those partners preserve citations, link-backs, and distinctions such as retractions versus versions of record. “We didn’t just agree with anybody; we need partners to have that appreciation and respect for the content,” Jarrett said. He named the OpenEvidence deal alongside a Claude connector with Anthropic, a Mistral connector and an AWS partnership as examples of the approach.
The Anthropic partnership, in particular, goes beyond content licensing. Jarrett said the two organizations have agreed to collaborate on a set of principles for how AI agents and systems should interpret scientific data. The alliance spans how to signal the difference between a version of record and a preprint, what a retraction means, and the fact that scientific articles published over time represent an ongoing dialectic, not settled truth.
Customers want a different relationship
Jarrett said Wiley is also seeing a shift in how long-standing institutional customers engage with its content. Subscribers who have held licenses for decades are coming back asking for something fundamentally different: API-level access so their AI agents can query Wiley’s corpus directly in real time. “They’re saying, ‘We want our agents to be able to talk to this corpus of knowledge via an API,'” he said. “‘How do we make a real-time feed so those agents can work on behalf of the research analyst, or all these new user profiles that would never have looked into this content before?'”
He described the trend as a “democratization of scientific knowledge:” AI is making specialized research accessible to people who would never have navigated a journal article on their own, but only if the underlying sources are trustworthy.
The novice-expert gap
One risk that Jarrett flagged is that out-of-the-box AI tools may be most dangerous for users in the middle of the expertise spectrum. “A lot of these out-of-the-box AI results are really powerful if you’re a novice or an expert,” he said. “If you’re just trying to get the basics, great. If you’re an expert, you can spot where it’s off. But if you’re in the middle, that’s where it’s harder to interpret the results.” He argued that high-stakes, R&D-intensive use cases, including clinical decisions, drug discovery and civil engineering, are where vertical-specific tools built on trusted content will outperform general-purpose LLMs, “at least for some period of time.”
