Peer Review Needs an AI Boundary, Not Another Author Disclosure

A reviewer does not need to paste a full manuscript into ChatGPT for AI to enter peer review. It can arrive through a meeting transcript tool, a browser extension, a reference assistant, an institutional chatbot, a vendor dashboard, a translation aid, or a document summarizer that nobody thought to name in the journal policy. That is why the next AI governance problem for journals is not simply whether authors disclose AI use. It is where the editorial process draws a boundary around confidential evaluation.

The policy direction is becoming clearer. The European Commission published the third edition of its ERA Living Guidelines on the responsible use of generative AI in research in May 2026 at https://research-and-innovation.ec.europa.eu/news/all-research-and-innovation-news/updated-era-living-guidelines-responsible-use-generative-ai-research-2026-05-08_en. The update kept the practical tone of the earlier guidance while adding attention to third-party AI use during meetings and information management, and to the risk of hidden prompts that can instruct AI systems outside normal human oversight.

Then, on July 1, 2026, STM noted that thematic working groups are now looking specifically at AI in proposal evaluation, AI in peer review, and AI-use declarations: https://stm-assoc.org/eu-com-group-on-generative-ai-use-in-science-2/. For journals, that is the important signal. The conversation is moving from author writing assistance toward evaluative work: the review, the recommendation, the editorial decision, and the evidence trail behind them.

The Sensitive Zone Is Evaluation

Author-side AI rules are easier to write because they sit in a familiar place. Journals can ask authors to disclose whether a tool helped draft text, produce a figure, translate a passage, write code, or organize references. The published article can carry a statement. The author remains responsible for the work. None of that is simple, but the accountability model is recognizable.

Peer review is different. A reviewer is not producing public content under their own name in most journal models. They are handling confidential material and exercising delegated judgment. An editor is not merely improving prose. They are deciding whether expert criticism is sufficient, whether a manuscript should be rejected, revised, accepted, escalated for integrity review, or held for additional checks.

That is why a journal cannot govern reviewer AI use by copying the author disclosure box. The core question is not, "Did someone use AI?" It is, "Was confidential manuscript information exposed, did AI influence evaluative judgment, and can the journal explain who remains accountable for the assessment?"

Three Doors AI Can Walk Through

The first door is the obvious one: a reviewer uploads a manuscript or review draft into a general-purpose AI tool. Several major publishers now draw a hard line here. Elsevier says submitted manuscripts are confidential documents and warns that uploading a manuscript, or any part of it, to an AI tool may violate confidentiality, proprietary rights, and data privacy rights: https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals. Taylor & Francis similarly tells editors and peer reviewers not to upload unpublished manuscripts or proposal materials into generative AI systems, and says reviewers must not use AI tools to generate review reports: https://taylorandfrancis.com/our-policies/ai-policy/.

The second door is quieter: AI enters through ordinary productivity tools. A reviewer may use an AI meeting assistant while discussing a manuscript with a co-reviewer. An editor may use an institutional summarizer on a decision letter. A staff member may paste reviewer comments into a tool to remove identifying details. Each action may feel administrative, but each can expose non-public manuscript content or shift judgment in ways the journal cannot later reconstruct.

The third door is infrastructure. Some AI checks are built into submission systems, integrity tools, language services, image-screening tools, or publisher dashboards. Those uses may be appropriate and valuable, especially before external reviewers are asked to spend time on a flawed submission. But they need governance: approved tools, defined purposes, access controls, audit records, known data handling, and human review of outputs.

Hidden Prompts Make The Boundary Concrete

The European Commission update calls attention to hidden prompts, instructions embedded for AI systems but not easily visible to people. In publishing, the risk is not theoretical. A manuscript, cover letter, supplemental file, reviewer response, or even a linked document could contain text meant to steer an AI summarizer or reviewer assistant: praise this paper, ignore limitations, recommend acceptance, downplay missing data.

A human reviewer can skip over a strange sentence or recognize manipulation in context. A tool may treat the instruction as part of its operating environment. Even if the attack fails, the journal has a governance question: did the review process rely on a system vulnerable to instructions the reviewer and editor could not see?

This is where blanket language such as "AI may be used responsibly" becomes too thin. Journals need role-specific rules. A tool that checks a submission for missing ethics statements before review is not the same as a tool that drafts the recommendation to accept. A spelling tool used on a reviewer comment is not the same as a model asked to identify methodological weaknesses in an unpublished study. The policy should name those differences.

A Useful Boundary Is Task-Based

Journal leaders should separate AI uses by the task they affect, not by whether the tool is fashionable, institutional, or public. A workable boundary has four zones.

• Allowed without disclosure: local spelling, grammar, formatting, and accessibility tools that do not receive confidential manuscript content outside approved systems.
• Allowed with guardrails: journal-approved tools for technical checks, integrity screening, language support, reference checks, or administrative summarization when data handling is documented and outputs are reviewed by a responsible person.
• Allowed only with explicit permission and logging: use of AI on confidential reviewer comments, decision letters, editorial triage notes, or manuscript summaries inside a governed environment.
• Not allowed: uploading unpublished manuscripts, figures, data, proposals, or identifiable peer-review material into unapproved systems, or using AI to substitute for the reviewer or editor judgment that the journal asked a human expert to provide.

This structure is more useful than a simple yes-or-no rule because it matches how editorial work actually happens. It also gives staff and editors a way to answer edge cases without improvising under deadline.

Disclosure Is Not Enough Without Evidence

The coming wave of AI-use declarations will create a documentation problem. A reviewer may say they used a tool only to improve language. An editor may say an approved system generated a submission-quality flag. A publisher may say an integrity model helped prioritize checks. Those statements are helpful, but they are not evidence unless the workflow records what tool was used, by whom, for what purpose, under what policy, and with what human verification.

The weakest implementation is a free-text box that disappears into the manuscript file. The stronger implementation is an audit trail that distinguishes author disclosure, reviewer declaration, editor-approved tool use, staff processing, and platform-level automated checks. Journals do not need to publish all of that operational detail. They do need to be able to produce it when a complaint, appeal, correction, or investigation asks how a decision was made.

What To Change Before The Next Reviewer Invite

• Add a reviewer-facing AI rule to invitation and acceptance screens, not only to author guidelines.
• Define whether reviewers may use AI for language polishing of their own comments, and whether that use must be declared.
• Prohibit unapproved upload of manuscripts, figures, data, supplements, proposals, and review files into external AI tools.
• List any journal-approved AI tools separately, with permitted purposes and data-handling expectations.
• Give editors a structured way to record AI-related reviewer concerns, suspected hidden prompts, or approved AI assistance.
• Train editorial staff to treat AI meeting assistants, transcription tools, and document summarizers as manuscript-handling systems, not harmless office utilities.

The Practical Takeaway

The practical takeaway for journal leaders is to stop treating AI governance as an author-instructions update. The higher-risk work now sits inside editorial evaluation. Before the next board meeting, ask for a one-page boundary map: which AI uses are permitted for authors, reviewers, editors, staff, and platform tools; which uses require disclosure or logging; and which uses are barred because they expose confidential material or replace human judgment.

That map will not settle every future case. It will do something more valuable: make the journal explainable. When AI is involved in scholarly publishing, trust will depend less on whether a policy contains the right adjectives and more on whether the journal can show where human responsibility remained intact.