How widely do peer reviewers actually use AI tools?

Frontiers reports that 53% of reviewers now use AI tools in peer review. The figure comes from its whitepaper Unlocking AI's untapped potential: responsible innovation in research and publishing, which draws on a global survey of 1,645 active researchers conducted in May and June 2025. The excerpt describes AI as having "rapidly become part of everyday peer review," with uptake higher among early-career researchers (87%), in China (77%) and in Africa (66%). Most reviewers, the excerpt says, use AI for drafting reports or summarizing findings, while the whitepaper sees "significant untapped potential" for rigor, reproducibility and methodological insight. The survey is Frontiers' own, and Frontiers is also the author of the policy recommendations that follow, so these are its measurements of its own field, not an independent count.

The excerpt's mechanism is a gap between practice and policy. Researchers already use AI, and "many express a desire for clear, consistent policy recommendations that would enable more advanced use." Frontiers answers with recommendations for publishers, institutions, funders and tool developers, framed as "a practical roadmap for aligning publishing policy with how researchers are already using AI." The levers it names are governance, transparency and training, with publishers, institutions and policymakers asked to build sector-wide policy together. The argument is that practice has moved ahead of the rules, and that written policy should follow the practice.

Set against the nearest notes, the claim gets tested from the other side. Does banning LLM use in peer review change review outcomes? reports a randomized experiment in which the rules a conference actually imposed barely moved scores and were widely broken. Frontiers reports what reviewers do and asks for clearer rules; the experiment suggests that rules alone may not change outcomes. The survey shares a genre with How soon do AI researchers expect artificial general intelligence?, another survey of a research community, but it asks about workflow and trust rather than forecasts. On the upside the whitepaper points to, Can inference scaling help reviewers catch errors humans miss? is the contrasting case: PAT's rigor result is a measured outcome, while Frontiers' "untapped potential" for rigor is an expectation the survey does not test.

The excerpt does not establish several things. It says "53% of reviewers," but the sample is 1,645 active researchers, and no denominator is given. "Adoption" is never defined, so the 87% and 77% figures cannot be compared with the 53% without the whitepaper's definitions. The policy recommendations are cut off after "These include:", so what Frontiers actually proposes is not visible here. At the strength this evidence allows, the excerpt supports a narrower claim: self-reported AI use in review is already widespread, and researchers want clearer rules. It does not show that clearer rules would improve review.