Yes, in most cases. The ICMJE Recommendations require authors to disclose the use of AI-assisted technologies in manuscript preparation, and state that chatbots cannot be listed as authors. Journal policies vary in wording and placement, but they converge on the same rule: use is permitted, concealment is not.
What exactly does ICMJE require?
The ICMJE Recommendations set two requirements that between them define the mainstream position across medical publishing. Authors must disclose their use of AI-assisted technologies in the preparation of a manuscript, and a chatbot cannot be an author.
The second requirement follows from the first principle of authorship rather than from any view about AI. Authorship carries accountability: an author must be able to take responsibility for the accuracy and integrity of the work and to respond to questions about it after publication. A language model cannot be accountable to anyone, cannot approve a final version, and cannot answer a post-publication query. The bar excludes it automatically.
The first requirement is the one authors actually have to act on, and it is broader than most people assume. It is not limited to text generation. Using a model to summarise literature, to translate a draft, to rewrite a methods section for clarity, or to generate code that produced an analysis all fall within "AI-assisted technologies used in the preparation of the manuscript".
What is generally not treated as disclosable is routine language tooling that has been part of academic writing for years: spelling and grammar checkers, reference managers, and similar assistive software. The practical dividing line most journals draw is whether the tool generated content or merely corrected what you wrote. If a model produced sentences, paragraphs, code, or images that ended up in the submission, disclose it.
Because policies differ in detail, the reliable move is to read the target journal's instructions for authors before submitting rather than assuming a general rule covers you.
Where and how should disclosure be written?
Disclosure belongs wherever the journal says it belongs, and the location differs by venue. The common placements are a dedicated statement in the acknowledgments, a declaration in the methods section, or a field in the submission system itself. Some journals ask for all three in different forms.
A usable disclosure names the tool, the version where known, what it was used for, and who verified the output. The last element is the one authors most often omit and the one that carries the most weight, because it converts a confession into a statement about quality control.
Two examples, both short:
| Weak disclosure | Stronger disclosure |
|---|---|
| "AI was used in preparing this manuscript." | "GPT-4 was used to improve the readability of the Discussion section. All content was reviewed and edited by the authors, who take full responsibility for it." |
| "ChatGPT assisted with the literature review." | "ChatGPT was used to draft an initial summary of prior work. Every cited source was independently retrieved and verified by the authors; no citation was taken from the model's output." |
That second pair matters more than it looks. A peer-reviewed test published in Scientific Reports found that a majority of the bibliographic citations generated by GPT-3.5 were fabricated, that GPT-4 still produced a substantial share of fabricated or erroneous citations, and that even citations to real works frequently contained substantive errors. An author who used a model anywhere near their bibliography and did not verify each entry has a real problem in the manuscript, quite separate from the disclosure question.
Write the disclosure at the time you use the tool, not at submission. Reconstructing which sections a model touched three months later is how incomplete disclosures happen.
What happens if AI use is not disclosed?
Undisclosed AI use is handled as an ethics issue, not as a technical infraction, and journals have an established route for it. COPE guidance gives editors a documented route for handling concerns raised about a published article, authorship disputes, and post-publication corrections, and undisclosed AI use enters that same machinery.
The consequences scale with what the concealment affected. A cleaned-up discussion section that was not declared is a correction and an awkward exchange with an editor. Fabricated references that reached print are a different matter, because the published record now contains citations to work that does not exist. Fabricated data or images move into the territory of research misconduct as formally defined, which requires fabrication, falsification, or plagiarism committed intentionally, knowingly, or recklessly, per the US Office of Research Integrity.
There is also a detection dimension that authors underestimate. Editorial offices increasingly screen for AI-generated text and images as part of routine submission checks, and major publishers collaborate on shared screening infrastructure through the STM Integrity Hub. Generative tools leave statistical traces. The screening is imperfect and produces signals rather than proof, but the asymmetry is unforgiving: a disclosed use that a detector flags is a non-event, while an undisclosed use that a detector flags becomes a question about the author's candour rather than about the tool.
The reputational arithmetic is simple. Disclosure costs two sentences. Discovered concealment costs the credibility of everything else in the paper.
What should authors do now?
Treat AI disclosure as part of manuscript preparation rather than as an afterthought at submission, and the whole issue becomes administrative rather than fraught.
- Keep a running note of which tools touched which sections, updated as you write.
- Verify everything a model produced, especially references. Resolve every DOI to the actual record.
- Read the target journal's policy before submitting, because placement and wording requirements differ.
- Never list a model as an author, and do not credit one in a way that implies authorship.
- Disclose the borderline cases. Over-disclosure has no penalty; under-disclosure does.
The reason this discipline pays is that AI-related requirements are now one item in a longer list of things a journal expects a submission to satisfy, alongside reporting guidelines, data availability statements, and conflict declarations. Authors who are checking their work against the rules of wherever it is going tend to catch the disclosure question at the same time as everything else, rather than discovering it at the submission portal.
It also helps to separate two questions that often get merged. Whether you may use a model is a policy question, and the answer is usually yes with disclosure. Whether the output is correct is a quality question, and no policy answers it for you. A disclosed, verified use of a language model to tighten prose is unremarkable. An undisclosed, unverified use that introduced three references to papers nobody wrote is a problem that disclosure alone would not have fixed. Handle both, in that order. The quality half of that pair has its own article: how fabricated citations reach real papers.
How Octym helps
Octym reviews a manuscript for AI-generated text and image signals and for the declarations a target journal requires, and reports what it finds as evidence with its location in the document. These are signals for a person to weigh, never conclusions about an author's conduct, and every finding is traced to its source so it can be checked. You can see how the review is configured to a specific journal's requirements on the tailored page.