The race to put generative AI into scientific publishing has produced a policy scramble no one expected. Not over authors. Over referees.
Through 2023 and 2024, major journals and research funders quietly changed their rules about what a peer reviewer may do with a large language model. The U.S. National Institutes of Health, Elsevier, Springer Nature, Wiley, the JAMA Network and the Committee on Publication Ethics all published or updated guidance. The shared instruction: do not paste an unpublished manuscript into a public AI tool, and do not let one write the referee report.
It sounds like a narrow administrative fix. It isn't. Peer review remains the one part of journal publishing where confidentiality and expert judgment are the entire product. A weak paper can be corrected after publication. A leaked manuscript or a hallucinated critique cannot be recalled.
Generative AI systems such as ChatGPT, Claude and Gemini now sit inside everyday research workflows for drafting, coding, summarising grant text and preparing course materials. Publishers spent years encouraging those uses. A referee report is different. It requires a specific assessment of unpublished data, methods and interpretation under a confidentiality agreement. Once a manuscript enters a model with unclear retention settings, that bargain breaks.
Here's the catch
The rules are simple to publish and difficult to police. Editors see the final report, not the drafting process. A reviewer can run a private model on a local machine, paste a manuscript in sections, lean on an institutional tool with its own data terms, or ask a colleague to summarise the submission, and no editor will notice.
Automated detection doesn't close the gap. Text classifiers remain unreliable on short, technical, person-specific prose, and a false accusation against a referee is institutionally costly in a system that already struggles to find reviewers.
The policies also split on the most basic question: disclosure or prohibition. Some publishers allow generative AI if its use is declared. Others ban it outright. A few funders prohibit any AI assistance in grant review. A single reviewer may face three different standards in the same week.
That is the quiet failure inside the policy wave. The announcements demonstrate intent, but enforcement still depends on the honour system.
What the major players actually changed
In June 2023, the U.S. National Institutes of Health barred reviewers from using generative AI to analyse or critique grant applications, warning that such tools could compromise confidentiality and the integrity of the process. The NIH position is prohibition, not disclosure.
Elsevier's reviewer guidance tells referees not to use generative AI or AI-assisted tools to assist in the scientific review of a manuscript. The publisher's policy centres on confidentiality, personal accountability, originality and disclosure, and asks reviewers to keep submissions out of unapproved systems.
Springer Nature updated its editorial policies to say reviewers should not upload unpublished work to generative AI services and should not use the tools to draft their reports. If AI is used in an acceptable way, it must be disclosed to the editor.
The JAMA Network and journals published by the American Association for the Advancement of Science have issued comparable restrictions on entering manuscript text into AI systems. COPE, the Committee on Publication Ethics, keeps the burden on humans: the listed reviewer owns the report, and any AI involvement must be transparent without removing accountability.
Photo by Markus Winkler on Unsplash
What the prohibitions cover
Rules vary, but most new policies cluster around a small set of actions.
- Uploading an unpublished manuscript, abstract, figure or reviewer comment to a public generative AI tool.
- Using a language model's output as the basis of a referee report without reading and verifying the underlying paper.
- Delegating judgments about novelty, rigour, significance or reproducibility to software.
- Relying on AI-generated literature summaries that may fabricate or misattribute references.
The last point isn't hypothetical. Generative AI has a documented habit of inventing citations, and publishers worry those fabrications will migrate into referee reports. The earlier STM update on AI citation-stuffing in manuscripts described a similar pressure on the author side.
What usually remains allowed is the unglamorous work: grammar correction in a tool with no training access, basic reference formatting, summarising public information before the manuscript arrives and manual proofreading. The dividing line is less the technology than what the technology sees.
Hype versus reality
Reality grade: the bans are meaningful as a norm-setting exercise and weak as an enforcement mechanism. They will stop casual, unintentional leaks. They won't stop a determined referee with a local model or a university AI platform covered by a data-processing agreement.
Advocates for generative AI in peer review point to several legitimate uses. It can help a non-native English speaker polish a report. It can flag statistical inconsistencies at speed. It can summarise a reviewer's own notes. None of those require exposing the manuscript to a consumer chatbot.
The publishers know this. Their policies are not anti-technology so much as anti-convenience. The problem is that convenience is precisely what drives adoption, and journals have built no clear path for safe AI-assisted review at scale.
Indexation has never been a guarantee of review quality, as we've noted in reporting on journal indexation. The new AI rules are another reminder that policies on paper are only as strong as the editorial labour behind them.
Verdict: the policy wave is real and the enforcement gap is large. Clarity varies from journal to journal. Assume the strictest standard applies until an editor says otherwise.
What early-career referees should do
Peer review requests often arrive at the worst possible time for early-career researchers: during thesis writing, between experiments, in a job search and often without pay. The new rules add another layer. A careless upload can compromise a manuscript and your standing with an editor.
Before accepting a review, read the journal's generative AI policy. If you plan to use any AI tool, even for grammar, disclose it to the editor in a short note. If the policy is silent, ask.
Keep the manuscript offline. Use reference managers and statistical software that don't require uploading the submission. If you need language help, use a colleague or a paid human editor under confidentiality terms.
Editors are not your adversaries here. The worst outcome isn't being told no; it's being asked later whether a reviewer report that carries your name was written by software you didn't disclose.
For researchers applying to positions where peer review service counts, a documented record of careful, confidential review remains a credential. The new rules don't devalue that. They make the credential harder to fake.
Photo by Jonathan Kemper on Unsplash
The most pointed warning comes from the reviewers themselves. Every new policy shifts the burden sideways, not upward. A confidentiality rule only works if the person holding the manuscript refuses the shortcut. That is a professional norm senior researchers can teach, journals can reinforce, but no terms-of-use can install.