Using AI in a manuscript: disclosure, verification and responsibility

“AI was used” is too vague to explain what happened to a manuscript. Correcting grammar, drafting an introduction, generating analysis code and altering an image create different questions about reliability, disclosure and permission.

Make the unit of assessment the task, not just the tool’s name. Record what went into the tool, what came out, what was retained and how a person checked it. Then compare that record with the target journal’s current policy. This guide provides a working log and disclosure structure; it does not assume that all journals permit the same uses.

Check the policy for the task you performed

Read the journal’s author instructions and any publisher-wide AI policy they reference. Check whether the instructions distinguish writing assistance, research methods, software, figures and peer review. A rule for reviewers or book authors is not necessarily the rule for journal authors.

Record the policy URL, the date you checked it and any published version date. If the manuscript has been in preparation for a long time, check again before submission and after a transfer. Do not rely on an old screenshot or a colleague’s memory of what the publisher allowed last year.

ICMJE’s author guidance requires human responsibility and disclosure of relevant AI use, and does not permit AI tools to be listed as authors. Those recommendations apply in their disciplinary and adopting-journal context; they are not a substitute for reading the target journal’s instructions. 1

If the policy does not clearly cover your task, ask a precise question before using the output in a way that would be difficult to undo. Describe the task and affected material without sending confidential data merely to obtain a policy clarification.

Classify use by what it changes

Use the following distinctions when making the log. A single tool may appear in several rows because it performed several tasks.

Task	What the team needs to establish
Language editing or translation	Meaning, technical terms and degree of certainty remain accurate
Drafting or restructuring	Claims, interpretation and attribution are the authors’ responsibility
Literature discovery or summarisation	Sources exist and support the statements made
Code generation	Code performs the intended operation on the intended data
Analysis or modelling	Method, assumptions, inputs and outputs can be explained and checked
Figures or illustrations	Visual content is accurate, permitted and does not misrepresent evidence
Submission correspondence	Statements about the study and completed actions are true

Do not describe drafting as grammar correction because the final text was later edited. Do not describe code generation as language assistance because the tool also rewrote the methods section. The record should follow the actual contribution of the tool to the work.

This classification also helps decide who must review the output. A fluent writer may be able to check prose but not an unfamiliar statistical routine. The person verifying a task needs the relevant expertise, not simply access to the tool.

Language assistance still needs a meaning check

Compare edited or translated text with the original scientific meaning. Check negation, uncertainty, units, participant descriptions and causal language. A sentence can become smoother while becoming less accurate.

Pay attention to changes that strengthen a conclusion: “suggests” becoming “demonstrates”, an association becoming an effect, or a limited sample becoming a population. Ask whether the revised wording is supported by the actual design and results. Restore necessary qualifications even when they make the sentence less elegant.

Disclosure requirements can differ by task. PLOS’s current FAQ exempts language improvements such as spelling, grammar and rephrasing from its disclosure requirement, recommends disclosure of translation, and requires disclosure for code generation or development. That is a PLOS example, not a universal exemption. 3

Keep the use in your internal log even where public disclosure is not required. It gives the team an accurate account if a journal asks a more detailed question later. It also prevents different co-authors from making inconsistent statements about how the manuscript was prepared.

Verify references by reading the sources

Treat a generated reference as a lead until you have located the actual publication. Match the title, authors, journal or repository, year and identifier. Then read the relevant passage to establish whether it supports the sentence for which you intend to cite it.

A real DOI is not enough if it belongs to a different article. A real article is not enough if it addresses a different population, method or outcome. A correct summary is not enough if the cited work has a notice that changes its interpretation.

Elsevier’s current journal policy specifically requires authors to check AI-generated output, including sources that may be incorrect or fabricated, and to maintain human oversight. 2 Build that check into the writing process rather than leaving it for a final automated bibliography scan.

Use a claim table: manuscript sentence, source, supporting passage, limitation and person who checked it. If you cannot find the evidence, remove or qualify the claim. Do not keep asking a model to produce more plausible-looking references for a proposition that has not been established.

A literature summary should also preserve disagreement and scope. Check whether the tool omitted contrary findings, mixed a review’s interpretation with primary results or treated a tentative explanation as a settled conclusion. The author must decide what the evidence permits.

Code that runs is not necessarily correct

For generated code, write down the intended operation before reviewing the implementation. What are the inputs? Which observations should be included? How should missing values, duplicates, units and categories be handled? What outputs should be produced?

Use small, understandable test cases where you can determine the expected result independently. Check edge cases relevant to the real analysis, such as an empty group, missing observations or an unexpected category. Compare an important calculation with an independent method where practical.

Inspect dependencies, versions and any changes the tool made to the analysis plan. Keep the code that produced the reported results, not merely the final prompt. A rerun should use the documented inputs and environment, with a record of settings that affect the output.

If nobody on the team can explain or review the generated method, obtain qualified help before relying on it. A successful execution message does not establish that the method answers the research question. Do not conceal a tool’s substantive analytical role behind a generic statement about manuscript preparation.

Protect data before uploading it

Identify what the input contains: unpublished text, personal data, confidential reviewer comments, proprietary code, restricted datasets or third-party material. Check institutional rules, consent and contractual restrictions before uploading it to any external service.

Review the specific service and account settings rather than assuming all versions of a brand have the same terms. Elsevier’s author policy asks authors to examine privacy, confidentiality and rights in their tool use. 2 If the terms are unclear, use an approved institutional route or seek advice before sharing the material.

Public sharing of research data and uploading data to an AI service are different decisions. A journal’s data-availability requirement does not automatically authorise every processing destination. PLOS’s data policy, for example, explicitly addresses legal and ethical restrictions on access. 4

Use a minimal, non-sensitive input when that can accomplish the task. Remove identifiers only if doing so actually addresses the relevant risk; a distinctive combination of details can still be sensitive. Do not call data anonymous simply because names have been removed.

For a policy query or technical test, a synthetic blank structure may be enough. Label it as such and keep it separate from the real study records so that demonstration data cannot accidentally enter an analysis or figure.

Inspect figures as evidence, not decoration

Distinguish an illustrative diagram from a research image and a data-derived plot. Record whether the tool created shapes, wrote plotting code, altered pixels or generated the underlying visual content. Those actions should not be described interchangeably.

For a plot, verify the data-to-figure path: selected observations, transformations, axes, scale, labels, uncertainty displays and omitted values. For a schematic, check that its relationships and labels accurately represent the method or concept. For a research image, preserve the original and apply the journal’s specific image rules.

PLOS permits certain AI-assisted figure uses subject to accuracy, disclosure and its other policies, while prohibiting fabrication or misrepresentation of primary research. Do not convert that policy into a blanket statement that generated images are acceptable everywhere. 3

Also check permission to use input material. Creative Commons explains that a licence can grant only the rights the licensor controls and that other rights may remain relevant. An image being accessible online does not by itself resolve permission to upload, transform or republish it. 5

If the journal’s rule is uncertain, describe the actual workflow to the editorial office. “We used AI for a figure” may be too vague for an answer; “the tool generated plotting code from our specified structure, and we ran and checked it locally” is more precise if that is what happened.

Keep an AI-use log throughout the project

Complete the log when the work occurs. Reconstructing it from memory just before submission is less reliable, particularly when several collaborators used different tools.

Field	What to record
Tool and provider	Name, model/version if available, account or environment where relevant
Date and task	What the tool was asked to do
Inputs	Type of material supplied; permissions and sensitivity check
Output retained	Text, code, analysis, image or other contribution
Human changes	What was edited, rejected or rerun
Verification	Checks performed and the person responsible
Affected locations	Manuscript sections, figures, code or supporting files
Policy decision	Applicable journal instruction and disclosure location
Supporting record	Approved location for prompts, outputs, code and notes

Do not record secrets or sensitive data in a widely shared log. Refer to an appropriately controlled record where necessary. The purpose is traceability, not creating another uncontrolled copy of confidential material.

For a changing tool whose exact model version is unavailable, record what you can establish and say what you cannot. Do not invent a version number to make a disclosure look complete. Keep the task and verification description specific enough to remain useful.

Write a disclosure that describes the work

Use this as a starting structure, then follow the journal’s required wording and location:

During preparation of [manuscript or specified component], the authors used [tool, provider and available version] for [specific task]. The affected material was [sections, code, figures or analysis]. [Responsible authors or team] checked the output by [actual verification steps] and revised it as follows: [relevant changes]. The authors take responsibility for the submitted work.

For AI used as part of a research method, a short acknowledgement may not be sufficient. The methods description needs the information readers require to understand and evaluate the procedure. Determine what settings, inputs, model details and validation are necessary for that task, and follow the journal’s reporting expectations.

Do not claim that every output was independently verified unless that happened. Do not state that no confidential material was supplied if the team has not checked all contributors’ use. An honest disclosure may reveal an unresolved problem that must be addressed before submission.

Record rejected outputs too, when they matter

You do not need to preserve every trivial suggestion, but keep a note when an output exposed a consequential error or influenced a methodological choice. Record why it was rejected and what independent check resolved the issue. That can prevent another collaborator from accepting the same plausible but incorrect suggestion later.

Keep rejected draft text or demonstration code separate from the files used for the submitted analysis. Name the approved version clearly and make sure the manuscript describes that version. A complete disclosure is less useful if the team cannot identify which output actually entered the paper.

Check the manuscript and the declarations together

Before submission, ask each co-author about relevant tool use and reconcile the answers with the log. Check the manuscript, cover letter, submission-form questions and figure captions for consistency. A disclosure in one place does not necessarily satisfy a journal that asks for it elsewhere.

Revisit the policy if the manuscript is transferred or substantially revised. Keep a dated copy or note of the requirement used for the decision. If the policy and an editorial instruction conflict, ask for clarification through the official journal contact.

Before submission, ask the authors responsible for each part of the paper to explain how its claims, analyses and images were checked. Resolve any gaps before approving the final files, and retain the verification record alongside the disclosure.