How to report AI use in a research paper

Two uses of the same tool can create very different reporting duties. A researcher who asks ChatGPT to smooth three sentences in the introduction has changed the manuscript text. A researcher who asks an LLM to classify 2,000 survey responses has changed the research workflow, and maybe the results. Those cases should not be reported with the same sentence.

The difficulty is that journal policies have not settled into one universal format. ICMJE, Elsevier, JAMA Network, CDC, and Nature Portfolio all point in the same broad direction: disclose AI use, keep humans responsible, and give more detail when AI affects the study itself. The exact location and wording still depend on the journal.

Use the role of the tool as the sorting principle. AI that improves wording is usually a manuscript-preparation disclosure. AI that touches data collection, coding, analysis, image production, software, or results belongs in the research method.

What did the AI actually change?

The first reporting decision is whether AI changed the language of the paper or the evidence-generating process behind it. Writing assistance can often be described briefly, subject to the target journal’s instructions. AI used as a research instrument needs enough methodological detail for a reader to understand what was done and how the output was checked.

The table is only a starting point. Elsevier’s policy, for example, says basic spelling, grammar, and punctuation checks do not need its AI declaration, while substantive changes to sentence structure or organization do. It also says AI used as part of the research process should be described in Methods. ICMJE recommends acknowledgment for AI writing assistance and Methods reporting for AI used in data collection, analysis, or figure generation.

That is why copying a disclosure from another paper is risky. The other paper may have used AI for a different purpose, submitted to a different journal, or followed an older version of the policy.

Where should the disclosure go?

AI use can appear in several places: an acknowledgment, a dedicated AI declaration, the Methods section, a figure caption, supplementary material, the submission form, or the cover letter. The journal’s author instructions decide the final location.

The pattern across major policies is practical. If the tool helped write or revise the manuscript, the journal may want a short statement naming the tool, the purpose, and the human review. ICMJE says AI writing assistance should be reported in the acknowledgment section and disclosed at submission. Elsevier asks for a separate declaration for manuscript preparation uses covered by its policy.

When AI is part of the study, a short acknowledgment is not enough. The Methods section is where readers look for the materials and procedures that produced the result. An LLM that coded interview transcripts, generated variables, wrote analysis scripts, or produced experimental stimuli is part of that procedure.

A vague statement creates the wrong kind of transparency:

ChatGPT was used during manuscript preparation.

That sentence names a tool, but it does not say whether the tool corrected punctuation, rewrote the abstract, summarized sources, generated code, or classified participant responses. A useful disclosure names the action.

For writing assistance, the statement might say that an AI tool was used to improve readability in specified sections and that the authors reviewed and edited all changes. For research use, the statement belongs in a fuller method description.

What belongs in Methods when AI is part of the study?

Methods reporting should include the AI details that could change the result or that another researcher would need to understand the workflow. In studies using LLMs, JAMA Network guidance asks authors to report the platform or program, tool, version, manufacturer, dates of use, prompts, prompt sequence, and prompt revisions when those details apply.

Suppose an LLM classified open-ended survey responses into five categories. “We used ChatGPT to classify responses” leaves too many decisions hidden. A reader cannot tell which model was used, what the model saw, whether the prompt defined the categories, whether humans reviewed the output, or whether the authors reran uncertain cases.

Record these details while the study is running:

1. The tool or service used.
2. The model and version, when available.
3. The date or date range of use.
4. Whether access was through a website, API, local model, or institutional system.
5. The exact prompts or prompt templates that affected outputs.
6. Relevant settings, such as temperature or output constraints.
7. What data, documents, or variables were provided to the model.
8. How many runs were performed and how outputs were selected or combined.
9. How the outputs were checked, edited, tested, or adjudicated.

The date matters because online AI services change. Two papers can both say “ChatGPT” and mean different models, safety layers, retrieval settings, interfaces, and output behavior. If the method depends on the model response, a date is not clerical detail. It is part of the instrument description.

Prompts matter when changing the prompt could change the result. If a model assigned sentiment labels, extracted adverse events, generated synthetic cases, or answered questions that became data, the prompt is part of the method. Long prompts can go in supplementary material, an appendix, or a repository if the journal allows it.

How should repeated runs and model settings be reported?

Report settings and repeated runs when they could affect the output. For simple language editing, model temperature or sampling behavior rarely helps a reader evaluate the paper. For an experiment where model responses are the data, those settings can become the equivalent of an assay condition.

Generative systems can give different outputs for the same instruction. If you ran the model five times and selected the clearest answer, that selection is a methodological choice. If you asked the model to vote across runs, that is also a methodological choice. If a human coder resolved conflicts between AI labels and manual labels, the resolution rule belongs in Methods.

The reproducibility target is usually process transparency, not identical text. A future researcher may not be able to obtain the exact same sentence from an online model months later. They should still be able to see what model was used, what it received, what instructions it followed, which outputs were retained, and how the authors decided whether the output was acceptable.

For AI-assisted code, report the same things a reviewer would expect for hand-written code: what the code did, how it was tested, where it can be accessed if shared, and which outputs it produced. AI assistance does not relax the reproducibility standard. It adds one more provenance question.

How should authors check AI output?

Human review has to be described when it affects the final research record. ICMJE warns that AI output can be incorrect, incomplete, or biased, and it leaves responsibility with the human authors. CDC guidance also asks authors to report human oversight, including whether sources, extracted data, code, or results were independently verified.

Write down the checking procedure in terms a reader can audit. “All output was reviewed” is weaker than “two authors independently reviewed 10% of AI-coded responses, resolved disagreements by discussion, and manually reviewed every response assigned to the uncertainty category.” The second version gives the reader a process.

The same principle applies outside classification tasks. If AI-generated code was used, say whether tests were run and whether numeric outputs were reproduced independently. If AI summarized papers, say whether the original papers were read and checked against the summary. If AI suggested citations, cite the original evidence, not the generated statement.

An AI tool is never the source for a scientific claim. If it says prior studies found an effect, the paper still needs to locate those studies, read them, and cite the original source only if it supports the sentence. Fabricated references are the obvious failure. The quieter failure is a real paper attached to a claim it never made.

What should authors do about privacy and confidential data?

Do not upload sensitive, protected, confidential, or unpublished research material to a public AI service unless the service, institution, ethics approval, and consent terms allow that use. Disclosure after submission does not repair a privacy breach.

Check the inputs before checking the disclosure wording. Participant data, patient information, interview transcripts, unpublished results, collaborator documents, grant drafts, and copyrighted material can all create separate problems. Elsevier’s policy asks authors to consider data privacy, intellectual property, and the terms of the tool. CDC guidance says researchers should not enter sensitive, protected, or non-public data into public AI tools.

The useful questions are concrete:

• Does the service retain uploaded material?
• Can the provider use inputs for training or product improvement?
• Who can access the data?
• Does the institution permit this tool for this type of information?
• Did consent or ethics approval cover this processing?
• Are identifiers removed, and is removal enough for this dataset?

If the answer is unclear, pause before uploading. A carefully worded AI declaration tells the journal what happened. It does not make the use permissible.

Do figures and images follow the same rules?

Figures often have separate rules, and image policy can be stricter than writing policy. Elsevier distinguishes explanatory images, data visualizations, primary research images, graphical abstracts, and cover art. Under its policy, some explanatory images can be AI-assisted with disclosure, data visualizations must be tied to reproducible methods, and AI tools must not fabricate or alter primary research images.

A generated schematic may need a caption disclosure and a general AI declaration. A data visualization produced with AI support may need Methods detail about the data and workflow. A microscopy image, western blot, radiology scan, or patient image is a different category, because it represents observed or experimental evidence.

Nature Methods also emphasizes that policies around generative AI in scientific publishing continue to change. Treat figures as their own policy check before submission. Do not assume the writing disclosure rule covers visual material.

Can AI be listed as an author?

AI tools cannot be authors because authorship requires accountability. ICMJE says chatbots and other AI-assisted tools should not be listed as authors because they cannot take responsibility for accuracy, integrity, and originality. Elsevier and Nature Methods state the same principle in their publishing policies and commentary.

The practical consequence is simple. The human authors own the manuscript. They are responsible for checking AI-assisted text, code, figures, citations, and analysis, and they must be able to answer questions about any part of the work that used AI.

Acknowledging use is different from assigning authorship. A paper can disclose that an AI tool helped revise the language or classify text. It should not put the tool in the byline, cite it as if it were evidence, or shift responsibility to it when a generated output turns out to be wrong.

When does an AI reporting guideline help?

A reporting guideline helps when AI is central to the study rather than incidental to drafting. It should match the study type, field, and role of the model.

For medical research, the GAMER Statement provides a checklist for reporting generative-AI use across areas such as tool specifications, prompting, the role of the AI tool, content verification, privacy, and whether AI affected conclusions. TRIPOD-LLM is narrower: it is designed for healthcare studies that develop, tune, prompt, or evaluate large language models, with different items depending on the research design and LLM task.

Those guidelines are useful when they fit. They are not universal checklists for every paper whose authors used a chatbot to improve wording. A linguistics paper that used an LLM to generate experimental stimuli, a public health paper that used an LLM to extract data from reports, and a clinical paper evaluating an LLM system may need different reporting structures.

If a field-specific reporting guideline exists, use it alongside the journal’s author instructions. If no guideline fits, the journal policy still governs, and the Methods section still needs to explain the AI workflow clearly enough for readers to evaluate it.

What should be checked before submission?

Before submission, make an AI-use inventory rather than trying to remember one statement at the end. List every use of AI in drafting, editing, analysis, coding, literature work, figures, and supplementary material. Then sort each use by role.

Run this final check:

1. Check the target journal’s current AI policy and author instructions.
2. Separate writing assistance from research use.
3. Record tool names, models, versions, and dates where they matter.
4. Save prompts, prompt order, and settings that could affect results.
5. Describe how outputs were reviewed, tested, or adjudicated.
6. Verify every AI-suggested source against the original publication.
7. Check whether uploaded material was allowed under privacy, ethics, copyright, and institutional rules.
8. Put the disclosure in the section the journal requires.
9. Add Methods detail for AI that affected data, analysis, code, figures, or results.
10. Keep records in case the editor asks how the AI workflow was performed.

The standard scales with risk. A tool that corrected punctuation may need little or no disclosure, depending on the journal. A model that produced labels, code, images, or analytic outputs needs enough documentation for a reader to see what role it played and how the authors kept control of the result.

The strongest AI disclosure reports the work actually done. The closer the tool came to the scientific result, the more detail the paper should provide.

References and useful policies

1. International Committee of Medical Journal Editors. Use of AI by Authors.
2. International Committee of Medical Journal Editors. Defining the Role of Authors and Contributors.
3. Elsevier. Generative AI Policies for Journals.
4. Flanagin A, Pirracchio R, Khera R, et al. Reporting Use of AI in Research and Scholarly Publication: JAMA Network Guidance. JAMA. 2024;331(13):1096-1098.
5. Centers for Disease Control and Prevention. Considerations for Disclosing Generative AI Use in Scientific Work.
6. Nature Methods. Using AI Responsibly in Scientific Publishing. 2026;23:271.
7. Luo X, Tham YC, Giuffre M, et al. Reporting guideline for the use of Generative Artificial intelligence tools in MEdical Research: the GAMER Statement. BMJ Evidence-Based Medicine. 2025;30(6):390-400.
8. Gallifant J, Afshar M, Ameen S, et al. The TRIPOD-LLM reporting guideline for studies using large language models. Nature Medicine. 2025;31(1):60-69.