84% of researchers now use AI.

Use jumped. Expectations fell.

In one year, AI went from a minority tool to a majority habit in research. 84% of researchers now use AI in their work, up from 57% a year earlier, according to Wiley's survey of 2,430 researchers, fielded in August 2025. That is one of the fastest adoption curves in any profession the Good AI Prism tracks.

The same survey records a sharp reality check. Researchers believed AI outperformed humans on less than one third of the tasks presented, down from more than half the year before. Concern about hallucinations rose from 51% to 64%. And only 41% say their institution offers adequate support.

84% of researchers use AI in their work, up from 57% (Wiley, Oct 2025)

Read together, the numbers describe a field that has adopted AI faster than it has learned to govern it. Researchers use general assistants for literature, writing and code. Institutions have not yet settled which tools, which data and which checks they stand behind.

Source: Wiley ExplanAItions survey, 2,430 researchers, published 7 Oct 2025.

Science is where AI already won a Nobel

Science is also where AI has the strongest proof of value. The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design and to Demis Hassabis and John Jumper for AlphaFold, which predicts protein structures. A tool that took decades of methodological work became a standard instrument in biology within a few years.

The next stage is reaching the clinic. ETHORITY's Pharma, Biotech and Longevity study records that Phase 3 trials of drugs designed with generative AI have started: Generate in December 2025 and Insilico in September 2026. In cancer screening, the MASAI trial reported in The Lancet in 2026 that AI-supported breast screening cut screen-reading workload by 44%, with interval cancers non-inferior. Those are the outcomes research leaders can point to: fewer years to a candidate, fewer hours per read, the same or better safety.

Government research is a large share of public AI already. Science is the single largest topic among the 3,611 AI use cases US federal agencies reported for 2025, at 21.0%, according to the federal AI use-case inventory. The Public Sector topic area shows the rest of that portfolio.

Europe puts money on compute and talent

On 8 October 2025 the European Commission adopted a strategy for AI in science. Its centre is RAISE, the Resource for AI Science in Europe, launched as a pilot at the AI in Science Summit in Copenhagen on 3–4 November 2025 with €107 million under Horizon Europe, the Commission's strategy page states. €58 million of that funds networks of excellence and doctoral networks to train and keep AI and scientific talent.

€600m from Horizon Europe to widen researchers' access to compute, including AI Gigafactories (European Commission, Oct 2025)

A further €600 million from Horizon Europe widens access to computational power for science, including dedicated access to AI Gigafactories for EU researchers and start-ups. The stated aim is to double Horizon Europe's annual investment in AI to more than €3 billion. For a university or institute, that turns compute from a budget line into an application process, and data management into a condition for access.

Open data is the research advantage

Research runs on data that can be inspected. The Good AI Prism Registry documents the open corpora behind many models, each card with its licence, size and date. Common Crawl's registry card lists more than 300 billion web pages, released monthly. The Dolma dataset card records 2.3 trillion tokens of English text under ODC-BY, published by AllenAI for fully open model development. OpenWebMath's card holds 14.7 billion tokens of mathematical web text for reasoning research. Open-weight models such as gpt-oss-120b on its card run under Apache 2.0 on an institution's own hardware.

Licences differ more than most buyers expect. ImageNet's access agreement limits use to non-commercial research and education; Wikipedia's text is CC BY-SA. A model trained for a grant project may not be usable in a spin-off. The Good AI Prism Global Market Study proposes the measures that matter for science: time to publication and reproducibility, supported by open data and methods and documented model and data choices.

Support is the missing layer

The gap between 84% use and 41% support is where institutions can act. Researchers do not need another general chatbot licence. They need an approved set of tools for their discipline, clear rules on which data may leave the institution, and help to check what the tools produce. That is a service, not a purchase.

Three roles make the difference. A research-IT owner decides which assistants, coding tools and models are approved and where they run. A data steward decides which data classes may be processed by which tool, including patient data, unpublished results and data under third-party licences. And a methods lead, often in the library or a graduate school, teaches how to document AI use and check its output. None of these roles is new. What is new is that they now need to work together on every AI tool, before a lab adopts it on its own. Institutions that set this up early turn adoption into an advantage in grant applications, because funders increasingly ask how AI was used.

The rules give research room, not a free pass

The EU AI Act excludes AI systems developed and put into service for the sole purpose of scientific research and development, under Article 2(6) of the AI Act. The exemption ends where research ends. A model that leaves the lab, enters a product or supports decisions about people falls under the ordinary rules, including the high-risk duties from 2 December 2027 for uses listed in Annex III. Health uses add medical-device rules on top.

Research integrity rules do not wait for law. Journals, funders and universities increasingly ask authors to state how AI was used. Hallucinated references are the most visible failure, and 64% of researchers now worry about them. A documented model, prompt and data trail is the simplest defence.

In practice

Best practice: an AI statement in every manuscript. A lab records, for each paper, which model and version it used, for which step, with which prompts and data, and who checked the output. The statement goes into the methods section and the repository. It answers journals and funders, and it catches hallucinated references before a reviewer does.

Scenario: from grant project to spin-off. A group trains a model in a funded project under the research exemption of Article 2(6). Two years later it founds a spin-off to sell the model. At that point the exemption ends, and every dataset licence is checked again: research-only data such as ImageNet cannot follow the model into the product.

What to do Monday

1. Inventory research AI. List the assistants, coding tools and models your groups use, including personal accounts, and record where research data goes.
2. Choose documented models and data. Prefer models and datasets with stated licences, versions and training data, and keep the version used for each result.
3. Write a usage statement. Agree how AI use is disclosed in papers, theses and grant reports, and how citations are checked.
4. Plan compute access. Map which projects could apply for RAISE and EU compute, and what data management that requires.
5. Measure what science values. Track time to publication, reproducibility and error rates before and after a pilot, not only hours saved.