Choosing an AI disclosure posture journals will accept

Somewhere between “we used no AI tools” and “a language model drafted the introduction,” there is a sentence you will have to write in your next submission. Most authors write it at the very end, from memory, under deadline — which is how disclosure statements end up either vaguely over-broad or quietly incomplete. Both are avoidable. The journals have, over the past three years, converged on a reasonably stable set of principles, and a team that chooses its disclosure posture before drafting can write the statement in two minutes, from records, with nothing to fear from an editor’s follow-up question.

Where journal policy has landed

The early scramble of 2023-era policies has settled into common ground that nearly every major publisher now shares. Three principles do most of the work:

• AI tools cannot be authors. Authorship requires the ability to take responsibility for the work and approve the final version — software can do neither. Nature and the other Springer Nature journals, Science, the ICMJE recommendations, and COPE’s position statement all agree on this without qualification.
• Generative use must be disclosed. If a tool produced substantive content — prose, summaries of literature, analysis, figures — journals expect that documented, typically in the methods or acknowledgements. Accountability for the content remains entirely with the human authors: “the tool said so” has no standing.
• Assistive use is treated differently. Here the most influential precedent is Nature’s explicitly stated exemption: AI-assisted copy editing — improving the grammar, style, and clarity of text the authors themselves wrote — does not need to be declared. Many editors and publishers have adopted the same line in substance. The policy logic is that copy editing does not change who generated the ideas or the findings, so it does not touch what disclosure is meant to protect.

That third principle is the load-bearing one for everyday writing, because it draws the line every team needs: not “did you use AI?” but “did the tool generate content, or assist with content you generated?”

Assistive versus generative: the distinction that decides everything

The assistive–generative boundary sounds crisp and blurs quickly in practice, so it is worth stating carefully. A use is assistive when the intellectual content — the claims, the structure of the argument, the interpretation — existed before the tool touched it. Tightening a paragraph you wrote, fixing grammar in a second language, suggesting a clearer phrasing of your own sentence, checking whether your citations resolve: assistive. The tool’s removal would cost polish, not substance.

A use is generative when content originates with the tool: drafting a paragraph from bullet points, summarizing a literature you have not read, proposing an interpretation of results, producing a figure or its caption. The test we find most durable: if a reviewer challenged this passage, would your defense rest on your own reading and reasoning, or on the tool’s output? If any part of the answer is the tool’s output, treat the use as generative and disclose it.

Two edge cases deserve honesty. Translation of your own text sits near the assistive line but is worth disclosing when the target-language phrasing carries scientific nuance. And literature discovery is assistive only if you actually read what was found — a summary you cite without reading the source is generative content wearing an assistive costume.

Three postures that work

In practice, teams need a posture, not a philosophy — a rule simple enough that every co-author applies it identically. Three postures cover nearly everyone:

• Posture 1 — No AI involvement. Nothing to disclose, but increasingly worth stating affirmatively, since some journals now ask directly. Only claim it if it is true for every author; one co-author’s undisclosed drafting assistant makes the statement false for the whole author list.
• Posture 2 — Assistive only. Tools may polish author-written text, check references, and flag inconsistencies; they may not draft content. Under Nature’s copy-editing exemption and its equivalents, this posture usually requires no declaration — though a brief voluntary note costs nothing and reads as care, not confession. This is the posture we recommend as the default for most teams.
• Posture 3 — Generative with verification. Tools may draft or summarize, and every generated passage is verified by a named author before it enters the manuscript — claims checked against sources, citations resolved, reasoning owned. This posture is legitimate at most venues if disclosed specifically: which tools, for which sections, with what verification. Vague statements (“AI was used in preparing this manuscript”) attract editorial questions precisely because they could mean anything.

A disclosure statement that works: “Language-model assistance was used for copy editing of author-written text and for reference verification. No content, analysis, or interpretation was generated by AI tools. All references were resolved against Crossref and verified by the authors.”

One further note: the journal’s policy is not the only policy in play. Funders increasingly attach their own conditions — several now prohibit generative use in grant review while permitting disclosed use in proposals — and many institutions have research-integrity codes that bind their researchers regardless of venue. A posture chosen per manuscript should be checked against all three layers: the journal, the funder acknowledged in the paper, and the institutional code. In practice posture 2 clears all of them everywhere we have looked, which is a further argument for it as the team default; posture 3 is where the layers genuinely diverge and reading the fine print earns its time.

Write the statement from logs, not memory

The posture decides what you may do; the record proves what you did. Months pass between drafting and submission, co-authors work differently, and “reconstruct everyone’s tool use from memory” is not a process — it is a liability. The teams that handle this well keep a running log: which tool, which manuscript section, assistive or generative, who verified the output. At submission time the disclosure statement is a summary of the log, and if an editor or an integrity committee ever asks a follow-up, the answer is a lookup rather than an argument.

This is one of the places where tooling honestly matters. In Dissertatio, every AI interaction — what was read, what was suggested, what a human accepted — is recorded automatically and exportable as a per-manuscript AI usage report, so the disclosure statement is generated from evidence. The same design keeps the assistant on the right side of the postures above: it verifies, searches, and flags against real records; it does not write your paper, so posture 2 remains available to every author who wants it. But whatever the tool, the principle stands: log at the moment of use, disclose from the log.

A checklist before your next submission

• Read the AI policy of the target journal and its publisher — this month’s version.
• Choose a posture with all co-authors before drafting, and write it down.
• Apply the assistive–generative test to every tool in the team’s workflow.
• Log tool use per section as it happens; assign one author to own the log.
• Verify anything generative — claims to sources, citations to registries — before it enters the draft.
• Write the disclosure from the log, name tools and scope, and avoid boilerplate vagueness.

Disclosure is often framed as a compliance burden, but the framing undersells it. A precise disclosure statement is a credibility signal: it tells editors and reviewers that your team knows exactly how its manuscript was made. In a literature currently discovering fabricated references at uncomfortable rates, that knowledge — provable, logged, specific — is quietly becoming part of what “rigorous” means.