What AI Use Is Permitted and Must Be Disclosed in a Manuscript

Most researchers now use AI somewhere in a manuscript. Far fewer can say which uses a journal permits, which cross into misconduct, and which must be disclosed before the paper is ever submitted.

Every policy on artificial intelligence in scholarly writing rests on a single distinction: a tool assists a person who remains responsible, whereas an author is a person who can take responsibility. Generative systems fall on the tool side of that line, and no degree of sophistication moves them across it. This is the one point on which editorial and publication-ethics organizations across medicine, the sciences, and the humanities have converged, and almost every other rule in this area follows from it.

The reasoning is practical rather than philosophical. Authorship requires that a named individual approve the final version, vouch for the accuracy and integrity of the work, declare competing interests, and answer for the result if it is later questioned. A software system can do none of these things. It cannot be held accountable, cannot consent to publication, and cannot be contacted when a reader disputes a claim. For that reason a generative tool never appears in the author list, is never named as a corresponding author, and is not credited as a co-author of the ideas it helped to express.

The consequence for the author is the part that is easiest to overlook. Because the human is the author, the human owns everything the tool contributed. A fabricated citation inserted by a language model is the author's fabricated citation. A subtle misstatement introduced while a sentence was being smoothed is the author's misstatement. Delegating the drafting does not delegate the responsibility, and journals treat the two as inseparable. This principle, rather than any list of approved or forbidden tools, is what an author needs to hold in mind when deciding how far to rely on a system.

Within permitted use, most journals draw a further line between employing a tool to assist the author's own work and employing it to substitute for the author's judgment. Assistance is generally acceptable; substitution is not. The permitted category covers the uses that change how a manuscript is expressed or prepared without changing whose thinking it represents, and in practice these fall into two broad groups.

Using AI to sharpen your language

The most common and least contentious use is linguistic. Authors, and particularly those writing in a language other than their first, use generative tools to correct grammar, tighten wording, improve flow, and translate drafts. Because this affects expression rather than substance, it sits comfortably within what journals allow, and for many researchers it has replaced the paid language-editing services that served the same purpose. The condition attached to it is verification. A tool that rewrites a sentence can quietly alter its meaning, soften a claim the author intended to state firmly, or introduce an error that reads fluently enough to survive a careless proofread. The author is expected to confirm that the edited text still says exactly what the research found.

Letting AI help with code and literature

Beyond language, tools are increasingly used to scaffold analysis code, suggest relevant literature, summarize background material, or draft routine passages such as formatted reference lists. Many journals accept these uses when two conditions hold: the intellectual contribution remains the author's, and the author verifies every output before it enters the manuscript. A suggested reference is checked against the real source rather than trusted on sight; generated code is read and tested rather than assumed to work; a summarized background is confirmed against the papers it claims to describe. The recurring theme across all of these permitted uses is the same pairing — verification, and in many venues disclosure — that turns a convenient shortcut into legitimate practice rather than a liability.

The prohibited category is defined by substitution and fabrication rather than by any particular product. The clearest prohibition is authorship itself, already described, together with the closely related practice of citing a generative system as the source of a factual claim. Such systems produce fluent text that can include references, quotations, and details which do not exist, and a model is therefore never a reliable authority for a fact and never a citable source for one. A claim that matters must be traced to a real, verifiable document.

Between the plainly permitted and the plainly forbidden lies the use that most often causes trouble: letting a tool supply the thinking rather than the phrasing. Asking a system to interpret results, to decide what a finding means, or to generate the argument of a discussion section crosses from assisting the author into replacing the author, even when nothing is fabricated. The output may be defensible when read in isolation, yet it no longer represents the author's own reasoning, which is the very thing a reader is being asked to trust. Journals that permit drafting assistance still expect the interpretation, the claims, and the conclusions to originate with the author.

The graver prohibitions concern the integrity of the content itself. Inventing data, inventing sources, or presenting AI-generated images as authentic experimental results or genuine figures is misconduct of the same order as any other fabrication, and the involvement of a tool neither explains nor excuses it. The same applies to passing off substantial machine-generated text as original scholarship where the journal forbids it, and to concealing the extent to which a manuscript was generated rather than written. What is decisive in each case is not that a tool was used but that the record has been made to misrepresent what was actually done.

A quieter risk sits alongside these. Confidential or unpublished material — a co-author's unreleased data, a third party's manuscript, information shared under embargo — should not be entered into public tools that may retain or reuse it, because doing so can breach confidentiality before the work is ever published. The unifying principle behind every prohibition is straightforward: nothing may enter the published record as genuine that is not, and nothing may be hidden that a reader would need in order to judge the work fairly.

Disclosure is the mechanism that keeps permitted use legitimate. It converts a private convenience into a transparent part of the record, and its absence is precisely what turns an otherwise acceptable use into a problem when it later comes to light. The requirement is not onerous, but it is specific, and vagueness undermines it.

What an AI disclosure must name

An effective disclosure names the tool that was used and, where possible, its version; it states what the tool was used for, distinguishing for example between language editing of the whole manuscript, the drafting of a particular section, and the generation of code for an analysis; and it makes clear that the author reviewed the output and takes responsibility for it. A blanket line noting only that artificial intelligence was involved communicates far less than a sentence that identifies the tool and its role, and editors and reviewers read the difference as a signal of how carefully the use was managed. Precision here protects the author rather than exposing them.

Where the AI statement belongs in a paper

Placement follows function. When a tool touched the research or the analysis, the disclosure generally belongs in the methods section, where the conduct of the study is documented. When it touched only the writing, it belongs in the acknowledgments or in a dedicated statement on the use of artificial intelligence, where the journal provides for one. It does not belong in the author list, and the tool is not entered among the references. Ordinary, ubiquitous aids — spell-checkers, reference managers, basic grammar correction — are usually treated as standard software and require no declaration; the threshold is whether the tool shaped content or substantial expression rather than merely tidied what was already there.

The core of these requirements is remarkably consistent across the field. The refusal of AI authorship, the prohibition on fabrication, and the insistence that the author remain accountable hold almost everywhere. Where journals diverge is in the detail of disclosure: what triggers it, how much it must say, and where it is recorded. Some titles ask for a positive statement in every submission, requiring authors to declare either the tools they used or that they used none. Others call for disclosure only above a threshold, treating minor language correction as ordinary and reserving the requirement for uses that shaped content.

The permitted uses vary in the same way. A journal in a computational field may take routine code assistance for granted, while a journal reporting clinical research treats any tool that touched the analysis with far greater caution. Publishers and preprint servers may impose their own standards on top of a specific journal's, and a field that has settled on one convention may sit beside another that has not. Because of this variation, a disclosure that satisfies one venue can fall short at another, and a use that passes without comment in one discipline can draw scrutiny in a neighboring one. The practical implication is that the governing document is always the target journal's own current policy rather than a general impression of what is customary.

The requirements are not an expression of suspicion toward the technology. They protect three things that scholarly publishing cannot function without. The first is accountability: a named person, answerable for the work, who can be asked to defend or correct it. The second is the integrity of the published record, which depends on readers and later researchers being able to trust that the data, the sources, and the results are real. The third is transparency, the principle that others should be able to understand how a piece of work was produced, including the part any tool played in producing it.

Disclosure serves the author as much as the reader. A declared, permitted use is documented good practice and closes the matter. An undisclosed one, discovered afterward, can read as concealment even when the underlying use was entirely harmless, and the damage then attaches to the concealment rather than to the tool. Because policies still differ between journals and are revised as the technology changes, the obligation that rests on the author is not to memorize a single universal standard but to consult the specific journal's current instructions before relying on any of this.

Meeting the standard is largely a matter of doing a few deliberate things before submission. Read the target journal's author guidelines and, specifically, its policy on artificial intelligence, since these differ between titles and are updated as practice settles. Establish plainly what a tool actually did in the manuscript, and write a disclosure precise enough that a stranger could tell the tool's contribution from the author's.

Then verify what the tool produced, item by item: every reference against its real source, every figure against the data beneath it, every claim against the evidence that is supposed to support it. Keep confidential and unpublished material out of public tools entirely. Approached this way, generative technology becomes what the rules assume it to be — an instrument used in preparing a manuscript, disclosed openly and owned completely by the person whose name the work carries. That, and only that, is the condition under which a journal accepts its use.