Do journal AI policies prohibit using an AI reviewer on your own manuscript?

Journal AI policies address two situations and treat them oppositely. Peer-review policies bar referees from uploading someone else’s confidential manuscript to an external AI service. Author policies permit AI assistance on your own unpublished work and require a disclosure statement instead. Establishing which of the two applies to you resolves most of the uncertainty around these tools.

The two are often printed on the same page, under the same heading, which is how they get conflated. The grammar gives the split away once you look for it: clauses addressed to authors are built on disclosure, and clauses addressed to referees are built on a bar against uploading. Elsevier, Wiley and PLOS each carry both halves in the same policy document — Elsevier’s reviewers “should not upload”, Wiley’s editors and reviewers “are not permitted to upload”, PLOS’s submission content “may not be uploaded” by handling editors or reviewers.

Which rule applies to you, by role

Your role at the moment you open the file decides the rule, not the tool you intend to use and not the file format. The same manuscript, the same model and the same prompt are permitted for the person who wrote the paper and prohibited for the person asked to referee it.

The row that surprises people is the last one. A response-to-reviewers letter is your own writing, but the reviewer reports you are answering were written in confidence for the editor, and the confidentiality that covers the manuscript covers them too. Draft the letter with the reports beside you rather than inside the prompt — see how to write a response to reviewers.

Common questions about who can use AI

Can I upload my own paper to ChatGPT to get feedback before I submit it?

Generally yes. Publisher policies for authors are built on disclosure, not prohibition — you may use AI on your own unpublished manuscript and describe the tool and its use at submission. The constraint that can actually stop you is narrower: your institution’s list of AI services approved for unpublished research data, which the publisher’s permission does not override.

I’m reviewing a paper for a journal — can I run it through an AI tool just to help me understand it faster?

No. The prohibition covers any use of the manuscript in an AI tool, not only uses aimed at producing your written review — Wiley’s policy explicitly extends the bar to “any parts of manuscripts including figures and tables,” which closes exactly the reasoning that a quick comprehension check is different from a full upload. The manuscript stays out of the prompt entirely.

Can I use AI to help write my response to the reviewers, or are their comments off-limits too?

Your response letter is your own writing, so AI assistance on the letter itself follows the author rules, with disclosure. The reviewer reports you are responding to are a different document: they were written in confidence for the editor, and that confidentiality carries over to you. Draft your response with the reports open beside you, not pasted into the prompt.

Why the referee restriction is strict

A manuscript sent to a referee is confidential material belonging to someone else, shared under an expectation that it will not be distributed further.

Uploading it to a third-party service is a distribution the author did not agree to, which is why publishers treat it as a breach regardless of that service’s privacy terms. The restriction is about consent and custody rather than about the technology, and a better privacy policy does not cure it. If you are an assigned referee, assume the prohibition applies.

The primary wording is uniform across publishers. Elsevier’s generative AI policy for journals states that “Reviewers should not upload a submitted manuscript or any part of it into an AI tool as this may violate the authors’ confidentiality and proprietary rights,” and applies the identical sentence to editors. Wiley’s ethics guidelines are more explicit about scope: editors and reviewers are “not permitted to upload manuscripts (or any parts of manuscripts including figures and tables) into AI Technology.” PLOS writes that “submission content may not be uploaded by handling editors or reviewers to generative AI platforms, websites, databases, or other tools; this is essential for protecting the confidentiality of unpublished work.”

Note what Wiley’s parenthesis rules out: a single figure, a single table, one paragraph of the methods. Referees who reason that a fragment is not the manuscript are reading a clause that was written to close exactly that gap.

A manuscript already posted as a preprint complicates this without dissolving it. The text is public, so the confidentiality argument for the content is weaker — but the review process is not the same object as the text. Your identity as the assigned referee, your assessment, and the fact that this editor gave you this manuscript to evaluate stay confidential under the same policies quoted above, and none of that changes because the words are already on a preprint server. Uploading the preprint alongside a prompt asking an AI tool to draft your evaluation still puts your review, not merely the text, into a third-party system.

The one policy that is conditional rather than absolute

The ICMJE Recommendations are the exception worth knowing, because the clause is conditional and permission-based rather than flat. Reviewers “must maintain the confidentiality of the manuscript as outlined above, which may prohibit the uploading of the manuscript to software or other AI technologies where confidentiality cannot be assured,” and “must request permission from the journal prior to using AI technology to facilitate their review.”

Read closely, that is a two-step test rather than a loophole: confidentiality has to be assured, and the journal has to agree in advance. In practice the second step ends the matter, because the journal you are refereeing for probably publishes with a large commercial publisher, and that publisher-level policy is the stricter of the two. Where a publisher policy and the ICMJE Recommendations differ, the publisher policy is the one the editor will enforce.

The first step is worth taking seriously rather than skipping past, because it is where “confidentiality cannot be assured” gets its content. A data-processing agreement a journal negotiates directly with a vendor, at an enterprise or API tier with contractual no-retention terms, is the shape of arrangement that could plausibly satisfy it. An individual reviewer’s personal, consumer-tier account is exactly what the clause is written to exclude, regardless of that vendor’s general privacy policy or no-training marketing.

ICMJE also states plainly why the caution exists beyond confidentiality: AI “can generate authoritative-sounding output that can be incorrect, incomplete, or biased.” That is a quality argument, and it is separable from the custody argument. Confusing the two is what produces the unproductive version of this debate, in which a referee argues that the model is accurate — which was never the objection. Can peer reviewers use AI on manuscripts works through what a journal-side policy that actually holds up looks like.

What a referee may still do

A referee may use AI on the review text they wrote themselves, with the manuscript absent from the prompt, and must say so. Wiley permits AI use to “improve the clarity or quality of the written feedback in a peer review report” and requires the reviewer to “disclose that use to the handling editor.” PLOS requires that “any use of AI tools in peer review (e.g. for data assessment, translation, or language editing) must be clearly disclosed to authors in the review form.”

The accountability clause is the part referees skim. Wiley: “Responsibility for the accuracy, integrity, and overall content of a peer review report rests entirely with the individual reviewer and must not be delegated to AI Technology.” A report you did not verify line by line is your report regardless of what produced the sentences.

For multilingual referees this carve-out matters more than it looks. Polishing your own English in a report you composed is permitted under the policies quoted above; pasting the manuscript in to have the argument summarised for you is not. The distinction is what left your desk, not how good your English is. Multilingual scholars covers the same line on the authoring side.

Questions about what referees can still do

Does it matter which AI tool I use as a reviewer, or are they all treated the same?

They are treated the same. Elsevier, Wiley and PLOS write the referee prohibition against the act of uploading, not against a specific vendor, so a tool’s privacy terms or no-training claims do not change the analysis. The one arrangement that could differ is a confidentiality agreement the journal itself negotiates with a vendor — an individual reviewer’s personal account carries no such guarantee.

Can I use AI to polish my English in a peer review report I wrote myself?

Yes. Wiley permits AI use to improve the clarity or quality of the written feedback in a peer review report, provided you disclose that use to the handling editor. The line is what left your desk: polishing text you composed is permitted, pasting the manuscript in so the tool can summarise it for you is not, regardless of how good or limited your English is.

What if my institution blocks a tool but the journal’s policy allows it?

The stricter rule governs. Institutions may restrict which services can receive unpublished research data regardless of what a publisher permits, and that institutional restriction is usually the stricter one. Ask your research office for the current approved-service list rather than requesting a one-off exception — exception requests are rarely tracked, while approved lists are actively maintained and updated.

Why the author position is different

Your own unpublished manuscript is yours to process.

Publishers who have addressed author use generally ask for disclosure of what was used and for what purpose, hold the author responsible for the content regardless of the tools involved, and decline to list an AI system as an author. None of that amounts to a prohibition. Reading it as one denies authors the use of tools they are permitted to use, which is the most common practical consequence of the confusion.

One qualification specific to co-authored work is easy to miss: “your own manuscript” is not the corresponding author’s alone to decide about. Where a paper has co-authors, the choice to run a joint draft through an external tool belongs to all of them, not to whichever author happens to hold the file — a co-author who has not agreed to that has the same interest in the decision that an author has in keeping a confidential submission out of a referee’s hands. Confirm with co-authors before uploading a shared draft, not only before the disclosure statement gets written.

The ICMJE Recommendations section II.A.4 are the reference text, and the verb is the point: “At submission, the journal should require authors to disclose whether they used artificial intelligence (AI)-assisted technologies (such as Large Language Models [LLMs], chatbots, or image creators) in the production of submitted work.” The same section rules out AI authorship — “Chatbots (such as ChatGPT) should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work” — and closes with the allocation of responsibility: “humans are responsible for any submitted material that included the use of AI-assisted technologies.” COPE reached the same conclusion on authorship in its position statement Authorship and AI tools, published 13 February 2023.

The author-facing section is built on disclose. One page away, the reviewer-facing section is where confidentiality is said to prohibit uploading. That asymmetry is the whole answer to the question this page is named for.

Where the disclosure line falls

Disclosure is required for substantive assistance and waived for routine language correction, and the boundary is written down. Wiley states that AI tools used “solely for spelling, grammar, and general editing are not included in these disclosure requirements,” but that where AI has been used to “substantially edit, develop, or translate any part of a manuscript, its use must be described transparently.”

Four cases sit near the line, and each has a defensible answer:

• Grammar and spelling correction. Not disclosable under Wiley’s carve-out. This is the same category as a word processor’s checker.
• Translation of a full section. Disclosable. Wiley names translation explicitly, and a translated methods section is a text a reader may need to interpret carefully.
• Feedback on your draft that you then acted on yourself. Disclosable as assistance, with the authors’ verification stated. The output is advice; the revisions are yours.
• AI-generated figures or images. ICMJE’s disclosure clause names image creators alongside LLMs, and several journals go further and refuse generated images in the results entirely. Check the specific journal before generating anything that will appear as data.

One journal family draws a harder line than the rest. The JAMA Network instructions for authors state that use of AI, LLMs and chatbots to draft opinion manuscripts “is not permitted” — Viewpoints, Perspectives, A Piece of My Mind essays, poetry and Letters to the Editor. That is a content-type prohibition rather than a role prohibition, and it is invisible if you read only the general AI policy. Read the instructions for the article type you are submitting, not the site-wide page.

What a disclosure statement usually contains

A workable disclosure names the tool, states what it was used for, and confirms that the authors reviewed the output and take responsibility for the result.

A statement covering a manuscript review might record that an AI reviewing tool was used to obtain feedback on a draft, and that all revisions were made and verified by the authors. Check the target journal’s wording before submission, since some specify a particular section for it and a few ask for it in the cover letter instead.

PLOS is the most prescriptive of the publishers quoted here and is therefore the safest template to write to: contributions by AI tools “must be clearly reported in a dedicated section of the Methods, or in the Acknowledgements section for article types lacking a Methods section,” and the report must give “the name(s) of any tools used, a description of how the authors used the tool(s) and evaluated the validity of the tool’s outputs, and a clear statement of which aspects of the study, article contents, data, or supporting files were affected/generated by AI tool usage.”

Three elements, then: tool, use, verification. A statement that satisfies all three reads roughly like this.

The authors used PerfectPaper (version and date of use) to obtain structured reviewer-style feedback on a complete draft of the manuscript. No text was generated for inclusion in the manuscript. All resulting revisions were drafted and verified by the authors, who take full responsibility for the content of the submitted work.

Elsevier asks for the same substance in a different container — “a separate AI declaration statement included in their manuscript upon submission” — so the sentences above move between formats without rewriting. Placement is the only thing you need to confirm per journal, and the journal submission checklist is where to confirm it. If the journal asks for the statement in the covering letter instead, how to write a cover letter shows where it goes.

Frequently asked questions about disclosure

Do I have to disclose it if I only used AI to fix grammar and spelling?

No. Wiley states plainly that AI tools used “solely for spelling, grammar, and general editing are not included in these disclosure requirements” — the same category as a word processor’s built-in checker. The obligation starts where the tool did more than tidy your sentences: substantive editing, drafting, translation, or analysis of the content itself.

I already used an AI tool on my manuscript before I checked the policy — do I need to say something now?

Yes, and doing it now is straightforward. If you have not yet submitted, add the disclosure statement to the Methods or cover letter before you do — disclosing when no policy required it creates no downside. If the paper is already under review, write to the handling editor, state what was used and when, and offer the declaration text; voluntary disclosure reads as a correction, not a violation.

What exactly do I write in the AI disclosure statement?

Three elements: name the tool, state what it was used for, and confirm you reviewed and take responsibility for the result. PLOS asks for the fullest version — tool name, a description of the use, and how you evaluated the tool’s output — which is why writing to PLOS’s template is a safe default even for a journal with lighter requirements. Check placement per journal before submission.

When the obvious remedy is unavailable

Four situations come up repeatedly where the policy answer does not resolve the practical problem, and each has a next step.

The journal has no stated AI policy at all. Disclose anyway, in the Methods, using the three-element form above. A disclosure no policy required costs you nothing; an undisclosed use that a later policy or a reader treats as material is the risk you are managing. Default to the ICMJE Recommendations when the journal is silent, since most editors already work from them.

You are the referee and you want the help. You cannot upload the manuscript, and no privacy assurance changes that. The available moves are to ask the editor for permission in advance, which is the route ICMJE describes, or to decline the assignment if you cannot review it properly unaided, or to press the journal to run structured checks itself on submissions it already holds — the journal has custody and an author agreement, so a journal-operated check is a different act from a referee upload. That route is the one with capacity implications, covered in peer review capacity and desk screening at scale.

Your institution bars the service you wanted to use. The strictest applicable rule governs and the institutional one usually is the strictest, so the publisher permitting it does not help you. Ask the research office for the approved list rather than for an exception; approved-service lists are maintained and exceptions are not. The institutional AI procurement checklist is the document that gets a new service onto such a list.

You already used a tool and did not disclose it, and the paper is under review. Write to the handling editor, state what was used and when, and offer the declaration text. Undisclosed use disclosed voluntarily is a correction; undisclosed use discovered by someone else is a conduct question. The gap between those two outcomes is one email.

Where funders and institutions add requirements

Journal policy is not the only policy in play.

Funders increasingly address AI use in grant applications and reporting, and institutions may restrict which services can receive unpublished research data regardless of what a publisher permits. When these conflict, the strictest applicable rule governs, and the institutional one is usually the strictest. Is it safe to upload a manuscript to ChatGPT covers how to check your institution’s position, and the procurement checklist is what your research office will likely work from.

The funder rules are more absolute than the publisher rules, and they are dated. NIH notice NOT-OD-23-149, released 23 June 2023, “prohibits NIH scientific peer reviewers from using natural language processors, large language models, or other generative Artificial Intelligence (AI) technologies for analyzing and formulating peer review critiques for grant applications and R&D contract proposals,” and states that “uploading or sharing content or original concepts from an NIH grant application, contract proposal, or critique to online generative AI tools violates the NIH peer review confidentiality and integrity requirements.” NIH bars the analysis itself, not merely the upload — a stricter formulation than any publisher clause quoted above.

The NSF notice of 14 December 2023 sets the same asymmetry as the publishers: “NSF reviewers are prohibited from uploading any content from proposals, review information and related records to non-approved generative AI tools,” while “proposers are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used and how it was used to develop their proposal.” Prohibition for the reviewer, encouraged disclosure for the applicant. In Europe, the European Research Area Forum’s living guidelines on the responsible use of generative AI in research, published 20 March 2024, ask researchers to avoid generative AI in peer review and evaluation and to respect confidentiality and intellectual property when they do use it. If your work touches human-subject data, the data-sharing consents in play add a further layer — see ethics and data sharing.

Note what NIH’s notice does not reach. NOT-OD-23-149 binds the reviewers assessing a grant application — it prohibits them from using generative AI to analyse or formulate critiques of the applications in front of them — and it says nothing about the applicant who wrote that application. NIH has issued no parallel prohibition on an applicant’s own use of AI while drafting a proposal; the restriction is asymmetric by design, the same shape as NSF’s explicit reviewer-versus-proposer split above.

What editors say when disclosure is mishandled

Editors and referees rarely write the phrase “AI policy violation”. They name the specific defect instead, and the objections take a small number of recognisable forms, each of which maps back to a clause quoted above:

• A declaration that says a language model was used “for editing” without naming which sections were affected or what changed — the substantive-versus-routine boundary Wiley draws.
• Methods that do not state how the authors evaluated the validity of the tool’s output — the element PLOS requires by name.
• Cited references that cannot be located, with a request for DOIs.
• A declaration naming one tool where the acknowledgements name another.
• An AI system appearing in the author list, with a request to remove it and record the use in the Methods instead — the position ICMJE and COPE both take.
• A disclosure that appears only in the covering letter and must also appear in the submitted manuscript.

The reference objection is the one that escalates fastest, because it is checkable in minutes and it is the failure mode most associated with unverified generated text. Verify every citation you did not personally retrieve, and confirm that each one supports the sentence it is attached to rather than merely existing. Detection tools that claim to identify AI-written prose are not reliable enough to found an editorial finding, which is precisely why editors act on artefacts they can verify instead: non-existent references, citations that do not support the claim, and numbers that disagree between the text and the tables. Those are the same defects a structured pre-submission pass catches — see the overclaim check.

Frequently asked questions about funders and misconduct

Will NIH reject my grant application if I used an AI tool to help write it?

NIH’s notice, NOT-OD-23-149, restricts peer reviewers evaluating your application — it prohibits them from using generative AI to analyse or formulate critiques of grant applications — and says nothing about applicants using AI to help draft one. NSF’s parallel notice does address proposers directly, but only to encourage disclosure of AI use in the project description, not to prohibit it.

Is using an AI tool on my manuscript considered research misconduct?

Not on its own. Disclosed use of AI on your own manuscript, with the authors verifying the output, is ordinary practice under every publisher policy quoted above. What crosses into a conduct question is a specific failure: use that should have been disclosed and was not, or a referee uploading someone else’s confidential manuscript despite the prohibition. The tool itself is not the violation; the undisclosed or unauthorised use is.

Can I list the AI tool as a co-author since it helped so much?

No. ICMJE and COPE are explicit on this: an AI system cannot be an author because authorship requires accountability for the accuracy, integrity and originality of the work, and a chatbot cannot hold that accountability. However much drafting or feedback help it gave, the responsible parties are the named human authors, and the AI’s contribution belongs in the Methods or acknowledgements, described and disclosed.

Where PerfectPaper fits

PerfectPaper is an author-side tool. You use it on your own manuscript before submission, which places it under the disclosure regime rather than the referee prohibition.

Accounts are limited to verified academic researchers, and your content is never used to train a model — not the paper, not the model’s reasoning, not its reply, not the derived metadata. That addresses the question an institutional policy is most likely to ask, and how the guarantee is enforced explains what holds it up.

The boundary is worth stating as plainly as the permission. PerfectPaper is not a route around the referee prohibition: an assigned referee may not upload another author’s manuscript to PerfectPaper any more than to a general-purpose chat product, because the objection is the author’s lack of consent and no vendor can supply that consent on the author’s behalf. PerfectPaper reads .docx manuscripts and reviews original research articles, and it reports findings for a human author to accept, revise or reject — it does not decide anything and does not replace a referee. The wider position is set out in secure AI review for unpublished research, including what PerfectPaper does not claim.

Related

Can peer reviewers use AI on manuscripts · Is it safe to upload a manuscript to ChatGPT · Secure AI for unpublished research · Institutional AI procurement checklist · AI peer review tool

Review my manuscript