How to write an AI disclosure statement for a research paper

A good AI disclosure names the tool and model version, says what it did and to which parts of the paper, explains how you checked the output, and confirms that the authors take full responsibility. Put it where your journal asks: research use usually goes in the methods, writing help in the acknowledgements or a separate AI declaration, and figure help in the caption; ICMJE journals and the Science journals also want it in the cover letter. Whether grammar-only help counts depends on the publisher, as the table below (checked 9 October 2026) shows.

What must an AI disclosure statement include?

Publishers call it an AI disclosure statement, an AI use statement or a declaration of AI use. They word it differently but ask for the same facts. Wiley lists the tool name and version, the date or year of use, the tool's role, the sections affected, and the author's role in directing or reviewing it (Wiley AI guidelines). PLOS asks for the tools, how they were used, what they affected, and how you evaluated the validity of the outputs (PLOS). Taylor & Francis also asks you to confirm that you checked the tool's terms of use (Taylor & Francis Author Services).

A complete statement answers six questions:

• Which tool and model? The product and the model or version shown at the time.
• When? A month and year, or a date range.
• What did it do? Editing, translation, searching, summarising, coding, analysis, drafting or figures.
• Where? The sections, figures, scripts or datasets it touched.
• How was it checked? What a person verified, and how.
• Who is responsible? A sentence saying the authors reviewed the output and take full responsibility.

That last line is not a formality. ICMJE says humans are responsible for any submitted material made with AI-assisted technologies, because AI "can generate authoritative-sounding output that can be incorrect, incomplete, or biased" (ICMJE).

Where do you put it: methods, acknowledgements or cover letter?

The place follows the kind of use. Across the policies checked on 9 October 2026:

• AI used to do the research (data collection, analysis, code, modelling, systematic searching) goes in the methods, in enough detail to repeat it. ICMJE asks for "the tool used, version, and prompts where applicable" (ICMJE, Methods). COPE, Elsevier, Wiley, Taylor & Francis, PLOS, ACM and the Science journals agree; IEEE instead asks for all AI-generated content, code included, in the acknowledgements.
• AI used to help write (drafting, rephrasing, translation) goes in the acknowledgements under ICMJE, Wiley, IEEE and the Science journals. Elsevier wants a separate declaration section immediately before the references; Taylor & Francis wants a "Declaration of AI use" statement.
• AI used on figures is disclosed in the caption or legend (Elsevier, Wiley, Taylor & Francis, PLOS, Springer Nature); Elsevier, Taylor & Francis and Springer Nature also want it in the general statement.
• The cover letter: ICMJE asks authors to describe AI use "in both the cover letter and the submitted work", and the Science journals ask for it in the cover letter as well as the manuscript (Science editorial policies).

ICMJE also asks journals to require a yes-or-no answer at submission, so expect a question in the submission form. Keep it consistent with the paper.

A journal's own instructions for authors override its publisher's general policy. Taylor & Francis, for example, notes that some of its journals "may not allow use of Generative AI tools beyond language improvement" (Taylor & Francis AI policy).

Should I cite the AI tool as well?

Disclosure and citation are different things. Every policy below that addresses authorship says an AI tool cannot be an author, and ICMJE adds that "referencing AI-generated material as the primary source is not acceptable" (ICMJE, Use of AI by authors). If AI output is itself something you analyse or quote, cite it in your style guide's format; see How to cite ChatGPT, Claude and Gemini.

Do I need to disclose AI used only for editing or grammar?

Sometimes. This is where publishers differ most, as of 9 October 2026:

• No statement needed for spelling and grammar fixes at Elsevier, Wiley, PLOS and the Science journals. IEEE says disclosure is "not required, but recommended". ACM's current policy (updated 14 May 2026) no longer requires disclosure of any AI help with writing.
• Statement expected at Springer Nature, whose framework rates grammar and language editing as low risk but still says to "clearly describe use and confirm author accountability" (Springer Nature), and at Taylor & Francis, which asks authors to acknowledge "any use of Generative AI tools". ICMJE puts AI writing assistance in the acknowledgements without an exception for grammar.

The exemptions are narrow. Elsevier exempts basic checks of grammar, spelling and punctuation, but says that when a tool "makes substantive changes to sentence structure or organization of a part of the text, this should be disclosed" (Elsevier). Wiley requires disclosure when AI is used "to generate substantial text or restructure arguments". IEEE allows language improvement without acknowledgement only "if the tool does not change the meaning or add new intellectual content to the text" (IEEE PSPB Operations Manual). Asking a chatbot to rewrite a paragraph so it reads better is usually more than a grammar check.

Translation is usually treated as more than editing. Wiley and the Science journals require it to be declared, Taylor & Francis wants the declaration to name the author who checked the translation, and PLOS recommends disclosing it without requiring it.

When unsure, disclose. Wiley's guidance says to "err on the side of disclosure", and Springer Nature states that declaration "does not negatively influence editorial decisions".

What if I forgot to disclose AI use?

Fix it early. ICMJE says nondisclosure "may require corrective action and may be construed as misconduct in some circumstances". PLOS says undisclosed AI use may prompt an integrity investigation, delay peer review, or lead to rejection, retraction or an expression of concern. If the paper is still under review, tell the editor and add the statement. If it is published, ask the journal how it handles corrections.

AI disclosure policies compared (checked 9 October 2026)

Each row comes from the publisher's or organisation's own page, opened on 9 October 2026. "Not stated" means the pages checked do not address the point; it does not mean anything goes.

A few rules in these pages catch people out:

• Taylor & Francis does not allow generative AI "to create a first draft of a manuscript or sections of a manuscript".
• The Science journals' AI usage table needs no declaration for literature searches or gathering references, but does not allow entering all or part of the manuscript into an LLM or other AI tool when revising it (Science journals' guidelines for AI usage).
• Elsevier bars general-purpose AI image tools from graphical abstracts.
• Since 1 October 2026, arXiv limits submitters to two submissions per calendar month, citing a rise in thin and AI-written papers; its policy still permits AI in support of research "as long as it is disclosed" (arXiv blog).

AI disclosure statement examples and templates

These are templates, not any journal's official wording. Replace the brackets, delete what does not apply, then match the section title and phrasing your target journal asks for. Elsevier, for example, recommends its own sentence and section title.

Language editing only (acknowledgements or AI declaration; skip only where the journal exempts it):

"During the preparation of this manuscript, the authors used [tool], [model and version] ([developer]), in [month year] to correct grammar and improve the readability of [sections]. The tool did not add content, analyses or references. The authors reviewed every change and take full responsibility for the content of the article."

Literature search or summaries (methods if part of a review method; otherwise acknowledgements):

"[Author initials] used [tool, model and version] between [dates] to [identify candidate studies on / summarise papers about] [topic], using the prompts in [Supplementary File]. Every work cited was retrieved and read by the authors. AI-generated summaries were checked against the original papers and were not used as sources. [Number] suggested references that could not be found or did not support the statement were discarded."

Code or data analysis (methods):

"Code for [task, for example data cleaning and the models in Section 3] was written with assistance from [tool, model and version] ([developer]; [dates]). [Author initials] reviewed, tested and revised all AI-generated code and checked its outputs against [an independent calculation / known results / a second implementation]. The final code is available at [repository or DOI]."

Drafting text (acknowledgements or AI declaration; not allowed at Taylor & Francis for first drafts):

"[Tool, model and version] was used in [month year] to draft [section] from the authors' notes [and to suggest a reorganisation of the Discussion]. The authors rewrote the text, checked every factual statement and reference, and take full responsibility for the content."

Figures (caption, plus the general statement):

"Figure [n]. [Title]. Schematic drafted with [tool, version] ([month year]) from the authors' sketch and edited by the authors. It illustrates [concept] and does not present research data."

Never use generative AI to create or alter images that are data, such as microscopy, blots, scans or photographs of results. Elsevier, Wiley and Taylor & Francis prohibit it. The Science journals go further: even illustrative AI figures need the editor's agreement.

No AI used (Taylor & Francis's suggested wording): "The authors report generative AI was not used in their research or preparation of this manuscript."

Elsevier's "Declaration of generative AI" section

Elsevier asks for its own section, titled "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process", placed at the end of the manuscript immediately before the references. Many guides still quote the older title, "…in the writing process"; use the current one. Its recommended wording, as of 9 October 2026 (Elsevier, policy updated June 2026):

"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 published article."

How do I keep track of AI use while I work?

The statement is written at submission, often months after the work. By then the chat is buried, the model has a new name, and nobody remembers which paragraph began as AI text. Publishers expect notes. Elsevier says authors should document "the name of the AI tool used, the purpose of its use, and the extent of their oversight". Wiley says to keep documentation of all AI use, including whether it shaped key arguments or conclusions. Taylor & Francis requires full records, including prompts, when AI is part of the research method, to be shared on request.

A simple log next to the project is enough, one line per meaningful use:

A few habits make the log reliable:

• Copy the model name the interface shows at the time, not what you remember later. If it shows none, write the product name and the date, and say so in the statement.
• Export or save the conversation whenever its output reaches the paper, the data or the code. For research use, keep prompts and settings.
• Note what you rejected as well as what you kept. It shows that checking happened.
• Mention AI help in commit messages, so the code history carries it.
• Mark AI-drafted passages in your working document until they have been rewritten and checked.
• Keep the log with the data and code, and agree which co-author confirms the final statement.

A dated log is also useful if you later need to show what you had, and when; see How to protect a research idea before you publish.

Keeping the record as you go

Papex is a research record for work done with AI. Drafts stay private while you work. AI apps you connect, such as Claude and ChatGPT, or Claude Code, Codex and Cursor through the Papex MCP package, can, with your permission, prepare private drafts, propose changes and look up references. They cannot sign, accept their own proposals, make rights statements or publish; only you do that, and an AI app is never listed as an author.

The model and tool names an AI app reports are kept in the record as it reports them and labelled "Agent-reported"; Papex does not verify them. Papex receives only what the app sends in a tool call, not your conversation, so you still need your own log. The record adds which assistant reported doing what.

When you sign, the signature links you to one exact version. A later edit becomes a new version with its own signature and date; earlier versions stay as they were. Only a salted fingerprint of the signed version leaves Papex. It is timestamped with OpenTimestamps and anchored in Bitcoin, meaning a public Bitcoin block shows the fingerprint existed by that block's time. The proof (an .ots file) checks with any OpenTimestamps tool, without Papex or a wallet (more on what that shows).

The limits matter. Papex does not detect AI use, check results or replace your journal's statement. A signed record shows who signed which exact version and that it existed by a date; it is not peer review and does not show the work is correct or original. After signing you choose to keep a version private, share it with people you invite, or publish it.

See how AI apps connect to Papex at /docs/ai.

Sources

• ICMJE Recommendations (updated January 2026), checked 9 October 2026
• ICMJE: Defining the role of authors and contributors, checked 9 October 2026
• ICMJE: Use of AI by authors, checked 9 October 2026
• ICMJE: Preparing a manuscript for submission (Methods), checked 9 October 2026
• COPE position: Authorship and AI tools, checked 9 October 2026
• Nature Portfolio: Artificial Intelligence (AI) editorial policy, checked 9 October 2026
• Springer Nature: Editorial policies, checked 9 October 2026
• Springer Nature: AI use in manuscript preparation, checked 9 October 2026
• Springer Nature: AI use in research practice, checked 9 October 2026
• Science journals: Editorial policies, checked 9 October 2026
• Science journals' guidelines for AI usage (PDF), checked 9 October 2026
• Elsevier: Generative AI policies for journals, checked 9 October 2026
• Wiley: Ethics guidelines for authors, checked 9 October 2026
• Wiley: AI guidelines for researchers, checked 9 October 2026
• Taylor & Francis: AI policy, checked 9 October 2026
• Taylor & Francis Author Services: Using AI in your research and manuscript preparations, checked 9 October 2026
• PLOS: Ethical publishing practice (AI tools and technologies), checked 9 October 2026
• IEEE Author Center: Submission and peer review policies, checked 9 October 2026
• IEEE Author Center (conferences): Submission policies, checked 9 October 2026
• IEEE PSPB Operations Manual (amended 25 June 2026), checked 9 October 2026
• ACM Policy on Authorship (updated 14 May 2026), checked 9 October 2026
• ACM: Frequently asked questions on the authorship policy, checked 9 October 2026
• arXiv: Content moderation (policy for authors' use of generative AI language tools), checked 9 October 2026
• arXiv blog: Fair moderation, equitable access, and AI (1 October 2026), checked 9 October 2026