Disclosed, Not Banned: The Shape of Publishers’ Emerging GenAI Policies

Three years ago, a manuscript submission to a major academic journal asked the corresponding author for the usual things: corresponding author details, conflict-of-interest statements, ethics approvals, funding sources. Today, that same submission portal asks one more question. Did you use a generative AI tool in preparing this manuscript? If yes, which one, and what for?

The answer, as of September 2025, is a declaration appearing in the published paper itself, immediately above the references, with a standardised template that the publisher provides. The shift is small in its administrative form and quite large in what it represents about how the publishing system has decided to handle a technology that, by the usual rhythms of academic policy, should have taken a decade to address. It took roughly eighteen months.

Elsevier’s generative AI policy for journals, updated September 2025, is one of the clearest expressions of the line that major publishers have settled on. It is worth reading not because Elsevier is uniquely authoritative — the same shape now appears in some form across Springer Nature, Wiley, Taylor & Francis, and most major society publishers — but because the policy makes explicit what is implicit in the others, and because the choices it makes are the choices the rest of the industry has converged on.

A framework built on three audiences

The policy is structured around the three groups who actually touch a manuscript: the authors who write it, the reviewers who evaluate it, and the editors who decide its fate. The rules differ meaningfully across those groups, and the differences are themselves informative.

For authors, the policy permits the use of generative AI tools in manuscript preparation, on the condition of disclosure. The disclosure is specific: a separate statement in the manuscript, the name of the tool, the purpose of the use, and an explicit assumption of responsibility for the resulting content. Authors cannot be replaced by AI — AI cannot be listed as a co-author, because authorship implies responsibility and accountability that the policy notes can only be attributed to humans. Basic grammar and spelling checks need no disclosure; substantive use does. The policy also draws a firm line on images: generative AI may not be used to create or alter figures in submitted manuscripts, with the only exception being where the AI is part of the research methodology itself and is described as such in the methods section.

For reviewers, the policy is sharper. Reviewers must not upload a submitted manuscript or any part of it into a generative AI tool, on confidentiality grounds. The peer review process is treated as a confidential channel, and uploading the manuscript to a third-party service is treated as a breach of that confidentiality regardless of how convenient the tool would otherwise be. The confidentiality requirement extends to the review report itself; reviewers are explicitly asked not to use AI to clean up the language of a review they have already written, since the report may contain confidential information about the manuscript and its authors. Reviewers are also asked not to use AI to assist in the scientific review itself, on the grounds that the critical assessment is the work the reviewer is being asked to do.

For editors, the rules are essentially the reviewer rules with additional weight. Editors must not upload manuscripts to generative AI tools; editors must not use AI to make editorial decisions; the final decision remains an editorial one in the conventional sense. The publisher reserves the right to use its own AI tools internally — for completeness checks, plagiarism screening, reviewer identification — and those tools are described as operating under the RELX Responsible AI Principles, which the policy links to as a public commitment.

Three audiences

Where the line falls, and why

The policy permits AI where the work product is the author’s own contribution, refined by tools. It prohibits AI where the work product is supposed to be a human act of judgement, performed under conditions of trust.

• AuthorsPermitted with disclosure

  May use AI in manuscript preparation. Must declare the tool, the purpose, and accept responsibility. Cannot list AI as a co-author. Cannot use AI to generate or alter figures, except as a research method.

• ReviewersProhibited

  Must not upload the manuscript or any part of it to an AI tool, on confidentiality grounds. Must not use AI to assist with the scientific review. Confidentiality extends to the review report itself.

• EditorsProhibited

  Must not upload the manuscript to an AI tool. Must not use AI to make editorial decisions. Publisher in-house AI may be used for completeness checks and reviewer identification, under the RELX Responsible AI Principles.

What the policy is actually doing

Read together, the three audience rules describe a coherent philosophical position, even if the policy does not state it explicitly. The position is that generative AI is permitted where the work product is the author’s own contribution, refined by tools; it is not permitted where the work product is supposed to be an act of human judgement, performed under conditions of confidentiality. Authors prepare manuscripts that they take responsibility for; AI tools may help them prepare. Reviewers and editors render judgement on manuscripts under conditions of trust; AI tools cannot help them render that judgement, because the judgement itself is the thing the system needs from them.

This is a more sophisticated position than the early conversations in 2023 anticipated. Those conversations tended to oscillate between two extremes: the prohibitionist view that AI use in scholarship is a categorical violation, and the laissez-faire view that AI is just a better word processor and requires no regulation at all. The policy that has emerged is neither. It treats AI as a category of tool whose ethical handling depends on what role the tool is being asked to play, and on whether the role is one that the system is willing to delegate to an algorithm. The answer for authors is mostly yes, with disclosure. The answer for reviewers and editors is mostly no, on grounds of both confidentiality and the irreducibly human nature of the assessment.

The disclosure template, in practice

The recommended disclosure format is short and revealing. A statement at the end of the manuscript, immediately above the references, reads roughly: “During the preparation of this work the author(s) used [NAME OF TOOL / SERVICE] in order to [REASON]. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.”

The template is doing a lot of work in a small space. It identifies the tool, which makes the disclosure auditable. It names the purpose, which separates substantive use from trivial use. It requires that the authors reviewed and edited the output, which makes the resulting text the authors’ work in a meaningful sense. And it places responsibility squarely on the authors, which is where the policy locates the accountability that AI tools, by the same policy’s reasoning, cannot bear.

The format also signals something about how readers will see this in five years. AI-assisted manuscripts will not be a separate class of paper. They will be ordinary papers with a disclosure footnote, indistinguishable in citation behaviour or scholarly weight from the manuscripts that came before them. The disclosure normalises the use; the responsibility clause maintains the boundary that matters.

A reflective note

It is unusual, in the history of academic publishing, for a major regulatory shift to be implemented this fast and this coherently. The norms around digital figures took roughly fifteen years to settle. The norms around data deposition are still settling. The fact that GenAI norms have crystallised in roughly two years, and have crystallised around a single framework that all major publishers can sign on to, says something about how unmistakable the threat to traditional editorial integrity felt, and how quickly the industry concluded that an absolutist position was not going to be tenable.

What remains uncertain is whether the framework will hold as the tools change. The current policy treats generative AI as a clearly identifiable category of tool. The next generation of writing assistance, with AI features woven into ordinary word processors and reference managers, will not be as easy to disclose, because the boundary between “AI used” and “AI not used” will be increasingly blurred. The policy as currently written is robust to that shift in spirit — the question it really asks the author is whether the AI tool did substantive work that the author needs to take responsibility for — but the implementation will need to keep moving. The September 2025 update will not be the last update.

For now, the practical implication for the working researcher is simple. If a tool helped substantively with a manuscript, disclose it in the standard format; if it did not, do not. The policy is written to be easy to comply with for authors acting in good faith, and difficult to evade for authors who are not. The system has decided that this is what good faith looks like, and the working researcher has the benefit, for once, of a clear answer to a question that two years ago had no settled response at all.