A cross-disciplinary study published in Learned Publishing finds that 83% of high-impact-factor journals now carry an explicit policy on AI use in peer review, against 75% among middle-impact-factor titles. The study, by Zhongshi Wang and Mengyue Gong (Jiangnan University), analyzed policies from 439 high-IF and 363 middle-IF journals across disciplines using a grounded-theory approach, and was published online December 23, 2025, appearing in the journal’s January 2026 issue (Vol. 39, Issue 1).
What the study measured
Wang and Gong’s analysis (DOI: 10.1002/leap.2035) is one of the larger systematic reviews of journal-level AI peer-review policy to date by sample size: 802 journals in total, split between a high-impact-factor group (439 titles) and a middle-impact-factor group (363 titles). Rather than starting from a fixed coding scheme, the authors used grounded theory to let policy themes emerge from the actual text of the guidelines they collected, then compared adoption rates and stringency across the two impact tiers and across disciplines.
The headline numbers
The core finding is a gap by journal tier: 83% of high-IF journals now have some form of AI guidance for the peer-review process, compared with 75% of middle-IF journals. That gap held up across disciplines but was not uniform in kind — the study also found a second, sharper divide by subject area. Science, technology, and medicine (STM) journals impose stricter, more detailed restrictions on AI use in review, while humanities and social-science journals tend toward more lenient, general-principle policies.
What the policies actually say
Where policies exist, the study found a consistent center of gravity: publishers emphasize transparency, human oversight, and confining AI tools to auxiliary tasks — grammar checking, literature/reviewer discovery — rather than substantive evaluation of a manuscript’s science. The recurring concerns cited across the sampled policies were confidentiality risk (submitted manuscripts entered into third-party AI tools), accountability gaps (who is responsible for an AI-influenced judgment), and AI’s inability to substitute for the critical human judgment peer review is meant to provide. This lines up with the individual publisher policies CASRAI already tracks in detail — see ICMJE’s generative-AI policy, Nature Portfolio’s AI policy, and IEEE’s generative-AI guidance, all of which restrict AI to a similarly narrow, disclosed, human-supervised role rather than banning it outright or leaving it unaddressed.
Why this matters for research administrators and editorial offices
For institutions managing journal relationships, editorial boards, or their own AI-in-peer-review practices, the practical takeaway is that a written AI policy is now the norm rather than the exception at the top of the impact-factor distribution — but it is still not universal, and mid-tier journals lag by a real, measured margin (75% vs. 83%). An editorial office or author checking whether a specific target journal has a policy still needs to verify directly rather than assume one exists, particularly outside the highest-IF tier. CASRAI’s publisher AI-policy landscape guide and AI-in-peer-review publisher comparison track the specifics for individual major publishers; this study is the first cross-disciplinary evidence of how consistently that individual-publisher pattern holds across the wider journal population.
Where policy development goes from here
The authors frame the disciplinary split as a call for tailored rather than one-size-fits-all guidance, proposing collaborative frameworks for responsible AI integration built around accountability, transparency, and interdisciplinary policy development — explicitly aimed at closing the gap between STM’s stricter posture and the more lenient norms in humanities and social-science publishing. That framing is consistent with the direction of the standards-setting activity CASRAI already tracks, including ANSI/NISO Z39.106-2023‘s work toward a shared peer-review vocabulary that any future cross-publisher AI-policy standard would need to build on.
