Most journals now require authors to disclose the use of generative AI. Far fewer say what an adequate disclosure looks like. The result is a lot of statements that satisfy nobody: too vague to be checkable, or so elaborate that a reviewer starts wondering what else the model did.
What follows is a practical account of what the major policies actually require, wording for the common cases, and the two failure modes that reliably draw objections.
The one rule everyone agrees on
Whatever else varies, this does not: an AI system cannot be an author.
COPE, ICMJE, the major publishers, and effectively every journal that has issued a policy converge here, and the reasoning is the same in each case. Authorship entails accountability — the ability to approve the final version, to vouch for the integrity of the work, and to answer for it afterwards. A model can do none of these. Listing ChatGPT as a co-author, which briefly happened in early 2023, is now a straightforward desk-reject.
The corollary matters more than the rule: you are accountable for everything the model contributed. Disclosure does not distribute responsibility. It documents a process. If a model introduced an error into your manuscript, it is your error.
Where the disclosure goes
Placement varies, and getting it wrong is a common minor irritation at submission:
- Language editing and formatting assistance generally belongs in the acknowledgements.
- Anything touching the research itself — data processing, coding, analysis, literature screening — belongs in the methods, described with the same specificity as any other tool.
- Many publishers additionally require a declaration in the submission system or a dedicated statement section. Elsevier, Springer Nature, Taylor & Francis, and Wiley each have their own field. Check the guide for authors of the specific journal; there is no universal placement.
The distinction is simple in principle: if the model touched the evidence, it is methods. If it touched only the prose, it is acknowledgements.
Wording that works
Concrete templates for the cases that come up most. Adapt rather than paste, and name the actual model and version — “an AI tool” is not a description of anything.
Language editing only
During the preparation of this manuscript the author used Claude (Anthropic, model version [X], accessed [month year]) to improve the clarity and readability of the written text. The author reviewed and edited all output and takes full responsibility for the content of the publication.
This is the most common case and the one publishers have been most explicit about. Several publishers supply near-identical template wording; using theirs is entirely acceptable and slightly safer.
Code generation for analysis
Analysis scripts in R were drafted with the assistance of GitHub Copilot ([version], accessed [month year]) and subsequently reviewed, corrected, and tested by the author. All statistical procedures were verified against [named reference implementation / worked example], and the analysis code is available at [repository link].
The verification clause is what makes this credible. Stating that generated code was checked, and against what, is the substance of the disclosure — the fact of generation is almost incidental.
Literature screening support
Title and abstract screening was conducted independently by two human reviewers. [Model, version] was used as an additional screening pass to flag potentially relevant records missed by both reviewers; all flagged records were then assessed by a human reviewer against the pre-registered inclusion criteria. The model was not used to make inclusion or exclusion decisions.
The final sentence is doing the work. In evidence synthesis, the boundary reviewers care about is whether a model ever made a decision that entered the results without human adjudication.
Qualitative coding assistance
An initial codebook was developed inductively by the author from [n] transcripts. [Model, version] was subsequently used to suggest candidate codes for the remaining transcripts; all suggestions were reviewed against the transcripts by the author, and [n] were retained, [n] modified, and [n] rejected. Final coding decisions and all interpretation are the author's. Inter-coder agreement was calculated between the two human coders only.
This case attracts the most reviewer scrutiny, reasonably so. Reporting the retention and rejection counts converts an unverifiable claim of oversight into a described procedure.
No AI use
The author declares that no generative AI tools were used in the preparation of this manuscript.
Increasingly required as an explicit statement rather than an omission. If you did not use anything, say so.
The two failure modes
1. Vagueness that cannot be checked
“AI was used to assist with the preparation of this manuscript” is the archetype. It names no tool, no version, no task, and no verification, and it covers everything from a spellcheck to a fabricated results section. A reviewer reading it learns only that something happened.
The fix is mechanical: name the tool and version, name the task, name the verification. Three clauses.
2. Over-disclosure that invites the wrong question
The opposite failure is rarer but more damaging. A statement listing every interaction across drafting, structuring, analysis, and interpretation raises a question the author did not intend to raise: what, exactly, is the author's contribution here?
Disclosure describes a process; it is not a confession, and it should not read like one. Report what the model did in the terms you would use for any other instrument. You would not write three paragraphs on your relationship with SPSS.
The line that actually matters
Underneath the wording is a substantive question about which tasks can be delegated at all. Disclosure norms are converging; the underlying boundary is clearer than the debate suggests.
| Generally accepted with disclosure | Not acceptable at any level of disclosure |
|---|---|
|
|
The organising principle is verification. A model may help you produce something you then check. It may not produce something you present unchecked, because the presentation asserts that a human stands behind it.
Citations, specifically
One point deserves separating out, because it accounts for a large share of the retractions and corrections attributable to AI use so far.
Language models fabricate references that look correct. Plausible authors, plausible titles, plausible journals, plausible years, and sometimes a DOI with a valid checksum pointing at nothing or at something unrelated. They are not detectable by eye, because they are constructed to be exactly what a real reference looks like.
The only defence is mechanical verification of every reference: resolve the DOI, confirm the article exists, and confirm it says what you have claimed it says. This is tedious and it is not optional. A fabricated citation in a published paper is a correction at best; several is a retraction. Both are permanent parts of your record, and both are entirely preventable.
Before you submit
- Read the target journal's guide for authors on AI. Do not assume it matches the last journal's.
- Decide, for each use, whether it touched the evidence (methods) or only the prose (acknowledgements).
- Name the tool, the version, the task, and the verification.
- Confirm no AI system appears in the author list.
- Verify every reference independently of whatever produced it.
- Keep a record of what you used and how. If asked later, you want the answer available rather than reconstructed.
Policies in this area are moving quickly and the specifics will change. The underlying requirement is stable, and it long predates the technology: describe your method accurately enough that a competent reader could evaluate it, and stand behind every claim you publish.