How to disclose Claude use in academic writing

Claude helped with your paper. Now say exactly how.

Claude can help a researcher rewrite a dense paragraph, test an argument, debug analysis code, summarize reviewer comments, or turn notes into a first draft.

That range creates a disclosure problem.

A sentence like "We used AI" tells an editor almost nothing. A sentence like "We used Claude Sonnet 5 to improve the clarity of the introduction and to draft candidate figure captions, then checked and rewrote all output" tells the editor what happened.

That is the difference.

This guide explains how to disclose Claude use in an academic paper, thesis, preprint, or conference submission. It also shows how to turn that disclosure into an AI Usage Card that you can store with your manuscript files.

If you need the broader version first, read How to disclose ChatGPT usage in academic papers or Do I need to disclose AI usage in my paper?. The logic is similar. The tool details differ.

What counts as Claude use?

Disclose Claude when it shaped the manuscript, research workflow, analysis, code, figures, or interpretation in a way that a reader or editor would want to know.

You do not need a dramatic threshold. If Claude touched the work in a material way, record it.

Common cases include manuscript editing, abstract drafting, translation, literature summarization, reviewer response drafting, code generation, data cleaning, classification support, and exploratory analysis planning.

Claude currently appears across model families such as Haiku, Sonnet, Opus, and Fable, and Anthropic publishes system cards for released Claude models. Record the exact model name when you can, because "Claude" alone may point to different systems over time. (claude.com)

Write down the date too.

A disclosure that says "Claude Sonnet 5, accessed through claude.ai on July 2, 2026" gives more value than "Claude was used." It lets a reader understand the technical setting without turning your methods section into a software manual.

For a general record format, see [AI Usage Cards examples and templates](/ai-usage-cards-examples/).

Where should you put a Claude disclosure?

Use the place your target venue requests.

Some journals ask for a separate AI declaration. Some ask for a statement in the methods section. Some allow an acknowledgment. Some submission systems include a text box.

Elsevier asks authors who use generative AI in manuscript preparation to include a declaration before the references, with the tool name, purpose, and a statement that the authors reviewed the output and take responsibility. (elsevier.com)

Nature Portfolio says authors should document LLM use in the Methods section or another suitable section, while AI assisted copy editing of human generated text does not need declaration under its policy. (nature.com)

ICMJE says AI tools should not appear as authors, that humans remain responsible for submitted material, and that nondisclosure may lead to corrective action or count as misconduct in some cases. (icmje.org)

So check the journal policy first. Then use the same factual core in the right location.

If your journal gives no format, use this order:

Name the tool and provider.
Name the model if known.
State the task.
State the scope.
State human review.
State that authors take responsibility.

That order works because it answers the editor's first question: "What did the AI do?"

For venue planning, read AI transparency requirements for journal submissions and [[[AI disclosure](/ai-disclosure-in-systematic-reviews-and-meta-analyses/)](/how-to-disclose-ai-use-for-neurips-icml-and-acl-submissions/) policies by major journals](/ai-disclosure-policies-by-journal/).

A short Claude disclosure statement

Use a short statement when Claude helped with language, structure, or small drafting tasks.

\section*{Declaration of AI-assisted tools}
 
During the preparation of this manuscript, the authors used Claude Sonnet 5
(Anthropic, accessed July 2026) to improve the clarity of selected paragraphs
and to suggest alternative titles for subsections. The authors reviewed,
edited, and approved all text and take full responsibility for the content.

This version fits many journal declarations.

It does not overclaim. It does not hide the tool. It tells the editor that Claude did not run the study, interpret the results, or write the final paper alone.

If you used Claude only for grammar and spelling, your journal may not require disclosure. Some policies treat copy editing differently from generative drafting. Still, your lab may choose to record the use internally, especially for dissertations, funded work, or multi author projects.

For thesis use, see How to disclose AI usage in your thesis.

A longer Claude disclosure statement

Use a longer statement when Claude helped with more than wording.

Maybe you asked Claude to draft a literature summary. Maybe you used it to generate R code. Maybe you used it to compare reviewer comments and author responses.

Name each role.

\section*{Declaration of AI-assisted tools}
 
The authors used Claude Opus 4.8 (Anthropic, accessed July 2026) during
manuscript preparation. Claude supported three tasks: drafting candidate
summaries of related work from author-provided notes, suggesting revisions to
the discussion section for clarity, and generating draft Python code for
descriptive checks of the dataset.
 
The authors verified all citations against the original sources, rewrote the
AI-assisted text, inspected and tested the code, and made all analytical and
interpretive decisions. Claude did not select the research question, collect
data, determine the conclusions, or approve the final manuscript.

This version gives editors what they need. It separates assistance from authorship.

That matters because publishers and editorial groups tie authorship to human accountability. Claude cannot approve a manuscript, answer questions after publication, or accept responsibility for errors. Do not list Claude as a coauthor. Use a disclosure instead.

For the authorship side, read Can AI be a co-author on a research paper?.

If Claude helped with code

Code needs a more careful note than prose editing.

If Claude generated code for cleaning data, simulation, annotation, statistics, model training, or visualization, put the details near the methods, software, or data availability section. The reader should know which parts of the workflow involved AI assistance.

A useful code disclosure names the language, task, review process, and repository.

\section*{AI assistance in code development}
 
The authors used Claude Sonnet 5 (Anthropic, accessed July 2026) to draft
initial Python functions for data cleaning and diagnostic plots. The authors
reviewed each function, revised the code, tested the outputs against manually
checked cases, and archived the final scripts in the project repository.
Claude-generated code did not determine inclusion criteria or statistical
model choices.

Do not bury this in the acknowledgments if the code affects results.

If Claude helped write analysis code, the disclosure belongs close to the method that used the code. A reader who wants to reproduce your results should see the role of the tool where they read about the analysis.

For tool choice and record keeping, see How to choose and disclose AI tools for research.

If Claude helped with literature work

Literature support creates two risks.

First, AI tools can invent citations or misstate papers. Second, a disclosure can sound more confident than the workflow deserves.

Do not write "Claude conducted the literature review" unless you truly delegated the search process and your venue permits that workflow. Even then, say what databases you searched, what Claude saw, and how you verified sources.

A better statement looks like this:

\section*{AI-assisted literature preparation}
 
The authors used Claude Sonnet 5 (Anthropic, accessed July 2026) to organize
author-provided notes on prior studies and to suggest possible thematic
groupings for the related work section. The authors checked all references
against the original publications and wrote the final literature review.
Claude was not used as a source of bibliographic evidence.

That last sentence matters.

An LLM output is not a source. It can help you think, but your paper should cite books, articles, datasets, protocols, software, and other materials that readers can inspect.

If Claude handled confidential material

Think before you paste.

Unpublished manuscripts, peer review reports, interview transcripts, patient data, student data, proprietary datasets, and grant reviews may carry legal or ethical duties. A good disclosure cannot fix a confidentiality breach after the fact.

Anthropic states that its API data practices differ by feature and product. Its API documentation says retained data does not train models without express permission, while its consumer privacy materials describe cases where chats or coding sessions may support model improvement if a user opts in, if a conversation gets flagged for safety review, or if the user joins certain testing programs. (platform.claude.com)

That distinction matters for research teams.

If you used Claude through an institutional account, record the product route, not only the model. "Claude via university enterprise account" and "Claude Free in a personal browser session" may involve different data controls.

For peer review, the stakes rise. Many publishers restrict uploading manuscripts or review materials to outside AI tools because review files belong to the confidential editorial process. See AI disclosure in peer review before you use Claude to draft a review.

Make a Claude AI Usage Card

A manuscript disclosure is usually short. Your project record can hold more detail.

That is where an AI Usage Card helps. It gives you a structured record of the tool, model, purpose, prompts, output handling, human checks, and limits.

You can generate one at ai-cards.org. Then you can save it as a PDF, keep it with your manuscript files, or copy selected text into the journal declaration.

For example, a Claude card for a journal article might say:

"Claude Sonnet 5 was used on July 2, 2026, through claude.ai to revise selected manuscript paragraphs for clarity and to draft candidate responses to reviewer comments. The authors checked all factual claims, rewrote the final text, and retained responsibility for the manuscript."

That is more useful than a vague note in a lab notebook.

It also helps when a coauthor asks, two months later, "Did we use Claude on the response letter or only on the introduction?"

You will have an answer.

Use the same record in Overleaf

If you write in LaTeX, you can keep the disclosure near the manuscript from the start.

Add a small file named ai-disclosure.tex:

\section*{Declaration of AI-assisted tools}
 
The authors used Claude Sonnet 5 (Anthropic, accessed July 2026) to improve
the clarity of selected author-written paragraphs and to suggest shorter
section headings. The authors reviewed and edited all AI-assisted text and
take full responsibility for the final manuscript.

Then include it before the references or wherever your journal wants it:

\input{ai-disclosure}
 
\bibliographystyle{plain}
\bibliography{references}

If the journal later asks for a shorter version, you can edit one file instead of hunting through the manuscript.

For more LaTeX help, use the LaTeX tutorial for AI Usage Cards or the Overleaf guide.

A simple checklist before submission

Before you submit, answer these questions in plain language.

Did Claude change the manuscript text, code, analysis, figures, translations, or interpretation?

Which Claude model did you use?

When did you use it?

Did you use a personal account, institutional account, API, or another route?

Did you paste confidential, personal, or unpublished material?

Did you verify every factual claim, citation, output, and code result?

Does your target journal require a specific AI declaration?

If you cannot answer one of these, fix the record before submission.

The goal is not to confess wrongdoing. The goal is to let readers see how the work came together.

Final wording you can adapt

For light editing:

During the preparation of this manuscript, the authors used Claude Sonnet 5
(Anthropic, accessed July 2026) to improve grammar, clarity, and readability
of author-written text. The authors reviewed and edited all output and take
full responsibility for the final manuscript.

For drafting support:

During the preparation of this manuscript, the authors used Claude Opus 4.8
(Anthropic, accessed July 2026) to draft candidate revisions of selected
sections from author-provided notes. The authors rewrote the text, verified
all factual claims and references, and take full responsibility for the final
content.

For code support:

The authors used Claude Sonnet 5 (Anthropic, accessed July 2026) to draft
initial R scripts for data cleaning and visualization. The authors inspected,
tested, and revised the scripts before analysis. The final code appears in the
project repository.

Now make the record while the details still feel fresh. Generate an AI Usage Card at ai-cards.org, save it with your manuscript, and copy the right disclosure text into your submission.