AI now mediates a third of scholarly information discovery, study finds
A third of scholarly information discovery is now being mediated by artificial intelligence, with many researchers unaware of the extent to which AI is shaping their view of the scientific literature.
These findings come from the first phase of Taming the Crocodile, a major industry study by Kudos that combines surveys of 11,500 researchers and 200 librarians with an environmental scan of more than 300 sources. The Phase I report has now been delivered to sponsors and includes analysis, recommendations and a strategic roadmap for publishers, librarians and technology providers responding to the rapid growth of AI-driven discovery.
Among the report’s most significant findings is that one in 10 researchers now begin their search for information directly in AI tools, while a further 25 per cent start with general search engines where AI-generated answers are increasingly embedded. The shift is even more pronounced among early-career researchers and those working in corporate environments, where the proportion beginning with AI tools doubles.
The study also highlights widespread misunderstanding about the role AI now plays in scholarly discovery. According to the findings, 40 per cent of researchers believe that AI-generated overviews in search engines have been created or curated by humans, rising to 46 per cent among students. Librarians reported that users are increasingly requesting articles that do not exist, having been directed to hallucinated references generated by AI systems.
The report suggests that the rise of AI-mediated discovery may already be affecting traditional measures of engagement with scholarly content. Librarians reported seeing declines in usage of primary sources as discovery behaviour changes, with many expressing concern that subscription purchasing decisions continue to be heavily influenced by usage metrics.
At the same time, librarians indicated that they want publishers to take a more active role in ensuring that scholarly content is represented accurately within AI systems. The report found strong support for greater transparency around publishers’ engagement with AI and thoughtful approaches to licensing content for AI training and agentic querying.
One of the clearest recommendations to emerge from the research is the need for publishers to create structured, plain-language content designed specifically for AI discovery. The report notes that AI optimisation specialists consistently recommend the development of “answer-ready” content that explains the importance and relevance of research in clear language.
The Phase I report includes guidance on content optimisation for AI environments, while Phase II will focus on developing industry best practice for schema, metadata and mark-up to improve AI visibility and attribution.
Additional recommendations include methods for benchmarking AI visibility, educational materials for librarians and researchers, guidance on generative engine optimisation and AI optimisation for search engines, strategies for maintaining visibility in China, and approaches to repositioning publishers as curators and stewards of knowledge rather than simply distributors of content.
“The findings give us all such a strong foundation to help adapt our industry infrastructure,” said Emilie Delquié, Chief Product & Customer Success Officer at Silverchair, which sponsored the study.
“This is the first evidence base we have seen that connects how researchers actually search to what authors and publishers should do about it,” said Shane Rydquist, Associate Vice President, Delivery and Solutions at CACTUS Communications. “The finding that matters most is that researchers rely on AI but do not trust it, and that evidence and clear attribution are what earn both the click and the citation. We are already using it to shape how we help authors make their work discoverable and credible in an AI-mediated world.”
Dan Penny, Head of Market Intelligence at Springer Nature, added: “The fast-moving and unpredictable impact of AI on all aspects of scholarly communication makes it essential to stay ahead. We all need to act quickly, and responsibly, on the findings here to ensure that we continue to support researchers, the wider community, and the integrity of scholarly communication.”
The next phase of Taming the Crocodile will move from diagnosis to response, examining the impact of AI-driven discovery on publisher usage and revenues, observing researcher behaviour directly, and developing shared standards for the content, mark-up and machine-readable signals that scholarly communication stakeholders want AI systems to ingest, interpret and attribute.
Further information about the project and the full report is available from Kudos.