Almost every author who asks me this has already done it once and felt uneasy afterwards. The manuscript is finished, you know it has soft spots, you want to know what a hostile reader would say, and there is a tool that will tell you in minutes. Then the doubt arrives: is this allowed? Does it count as cheating? Will an editor find out?
Most of that anxiety comes from collapsing two situations into one phrase. There is you, running your own unpublished manuscript through a tool to find its weaknesses before an editor does. And there is a reviewer, invited by a journal, pasting somebody else’s confidential manuscript into a chat window. Those two share a technology and almost nothing else. The first is about your data and your authorship. The second is about a document that was never yours.
I review for journals and I also built a tool that reviews manuscripts, so I have an obvious interest and you should read me knowing it. What follows is that distinction, what to ask any tool before you upload an unpublished paper, when disclosure matters, and where the line sits between using a model to criticise your text and using it to write it. Where I do not know something I say so, and point you to your journal’s instructions for authors, which is the only source that binds you.
The first is about you as an author. You wrote a manuscript describing work you did, and you want a machine to tell you where it is weak before you send it anywhere. No confidential document of anyone else’s is involved. The questions here are real but narrow: what happens to your text once you upload it, whether your coauthors agreed, and whether the tool ends up writing the paper rather than commenting on it.
The second is about a reviewer. A journal invited a specific person to evaluate a specific manuscript, one belonging to authors who never chose the reviewer and have no idea where their file might end up. That is not a reviewer working efficiently. It is somebody else’s unpublished work moving to a third party, and that is a question about confidentiality, not about workflow.
Notice how little they have in common. Advice that uses one word for both ends up paranoid about the first or far too relaxed about the second.
Start from what you already do without a second thought. You send the draft to a colleague, present it in a lab meeting where three people you barely know are sitting in, post it as a preprint where the entire internet can read it. Seeking criticism on your own unpublished work is not a grey area, so the question is never whether you are allowed feedback. It is what you just did with the file, and that turns on two things with little to do with peer review.
A multi-author manuscript is not yours to upload wherever you like. If one of your five collaborators works somewhere with strict rules about where internal material goes, sending the full text to an external service without telling anyone is a collegial problem waiting to happen, and it costs one message to the group to avoid. The point has more force for work under embargo, tied to a patent, or built on data governed by a contract with a health service or a school, where the constraint comes not from publishing ethics but from an agreement you already signed.
A manuscript normally reports aggregated results, so uploading the text is a very different act from uploading a dataset with identifiers, open-ended responses, clinical notes or interview transcripts. Qualitative work is the sharp case, because verbatim quotations can identify a participant even with names removed. Anything at the level of the individual participant is a separate decision, answered by your ethics approval and your consent form rather than by advice about AI. If you promised participants their data would stay with the research team, that promise has no exception for convenient software.
Using a model to criticise your text is a form of reading. It produces an opinion you judge, accept, argue with or ignore, and every word in the submitted paper is still a word you chose and can defend. That is not different in kind from a colleague scribbling "your sample size does not support this claim" in the margin. Commentary is not the paper.
Using a model to produce your text is a different act, because it has contributed content rather than an opinion about content. That is where authorship gets touched and disclosure becomes a live question. The reason is not mystical: authorship carries accountability. Someone has to answer for every claim and stand behind it when a reviewer pushes. A tool cannot do that, and if you have not read a sentence closely enough to defend it, nobody can.
In practice nobody sits neatly on one side. There is a gradient, from fixing typos and polishing sentences you wrote, through translating your own paragraph, to generating a section from a prompt. The early steps are what a bilingual colleague has always done. The last is the model writing your paper.
Take any passage and ask three questions. Can you explain why every claim is true, from your data or a source you have read? Can you defend the sentence without going back to the tool? Would you be comfortable if the editor could see how the passage came to exist? Three yeses and you are fine, whatever software touched the text. A no on either of the first two means you have shipped a claim you do not own, which would be a problem with no AI anywhere near it. A no on the third is your instinct noticing something the rules have not caught up with.
Your unpublished manuscript is a valuable asset and you only get to release it once. Whatever tool you are considering, general assistant or specialised reviewer, these questions are worth two minutes on its terms and privacy pages. I am deliberately not telling you what any product answers, because those answers change and you should read them yourself. If they leave you uncertain, upload less: you do not need the funding numbers, the author list or the supplementary appendix to get useful criticism of an argument.
I am not going to tell you what your journal requires, because I would be guessing, and a confident guess about somebody else’s policy is what gets repeated until it becomes folklore. What I can give you is the principle underneath most reasonable positions, and where to find the answer that applies to you.
The principle is about what a reader needs in order to evaluate the work. Disclosure exists so nobody is misled about how a piece of research came to exist. Nobody is misled by grammar correction. Someone could reasonably be misled if a model drafted substantive content, produced text that reads as your reasoning, or generated material you did not independently verify. The further along that gradient you sit, the more clearly your use belongs in the record.
The place to check is the instructions for authors of the specific journal, and then the submission system, which sometimes asks a direct question during upload. Those two sources are the ones that count. Not what a competitor journal does, not what your supervisor believes, and not this page. If the instructions are silent and you are unsure, email the editorial office and ask.
Keep it factual rather than apologetic. Name the kind of tool, say what you used it for, say what you did not use it for, and state that you take responsibility for the final text, adapting the wording to whatever the journal asks. What you should not write is a vague sentence covering everything just in case. One that says less than the truth is a problem; one that implies more involvement than there was invites questions you did not need to attract.
Prior publication exists as a concept to stop the same work being published twice as though it were two contributions, and as the term is normally used it turns on whether the work was made publicly available in a citable, findable form. Submitting text into a private tool is not that. It is closer to emailing a draft to a colleague, which is not the situation the concept was built to catch. One caveat: if a service publishes what you submit, into a public feed or a shared gallery, your text has been made public and the analysis changes.
Originality is the other fear, and the real risk runs opposite to what people assume. The danger is not that a model takes your text. It is that a model gives you text containing things you cannot stand behind, through two failure modes that are structural rather than accidental.
A general-purpose model produces text that is likely given what came before, and a citation is one of the most predictable objects in academic prose: plausible authors, a plausible year, a plausible title assembled from words your paragraph is already about. Producing something that looks like a reference is easy; producing one that exists is a different task. So every reference in your manuscript has to be one you opened and read enough of to know it says what your sentence claims. A fabricated citation that reaches submission is a citation error with your name on it.
The second failure is that assistants built to be helpful in conversation are shaped by that goal, so asking one what it thinks of your paper signals what you want and agreement is the response that satisfies the request. The test of any pre-submission review is whether it told you something you did not want to hear.
Now suppose an editor has invited you to review someone else’s manuscript. That manuscript is not yours. The authors did not choose you, do not know your name in most cases, and cannot consent to anything you do with the file. It is work that is not yet published, which in a competitive field can mean years of somebody’s career in a document that has not established priority. Whatever the invitation said about confidentiality is worth rereading in the invitation itself, and in the journal’s reviewer guidelines, rather than remembering it approximately.
There is a second reason beyond confidentiality. The editor invited you. Not a tool, not your lab, not the average of the literature. They invited you for a competence they believe you have, and the report is meant to be your judgement. A report you did not form is a report the editor did not commission, and passing it off as yours misrepresents the one thing the system runs on: a named expert taking responsibility for an opinion.
A rule of thumb while the norms settle: do not put the manuscript, or recognisable parts of it, into any external system. Checking the grammar of your own report afterwards is a much smaller act, though even there keep it to language and keep the specifics out. When you do not know what is permitted, ask the handling editor, and ask before the report goes in rather than after.
Nobody expects you to review with your hands tied. Reading about an unfamiliar method, or asking general questions about a statistical approach so you can evaluate its use competently, are things you can do without the manuscript leaving your desk. The distinction is whether the confidential document travels, not whether you used software while thinking. And if a paper needs expertise you do not have, the honest move has always been to tell the editor.
None of this overrides what your journal or institution tells you. It is what I do, as someone who both submits and reviews.
When people ask whether this is ethical, they often mean whether it feels like an unfair advantage. The advantage in question is having your work read critically before it reaches an editor, and that has always existed, distributed by institution: a supervisor who reads carefully, a manuscript club, a budget for language editing. Others submit into silence. Getting hard criticism before submission is not the part that needs justifying. What needs care is whose document you are handling, what happens to it, and whether you can stand behind every sentence you send.
No single rule covers every journal, so the answer depends on where you submit, and the instructions for authors are the source that binds you. Seeking critical feedback on your own unpublished work has never been controversial in itself. What varies is whether, and how, a journal wants AI involvement declared, particularly when a tool contributed to the text rather than commenting on it.
Check the instructions for authors of the specific journal, and look at the submission form, which sometimes asks directly. As a principle, disclosure exists so nobody is misled about how the research came to exist: language correction is not usually what disclosure statements are for, while substantive drafting sits much closer to the centre of the concern. If the instructions are silent and you are unsure, emailing the editorial office costs one paragraph and settles it.
That is the wrong question to optimise for. If your use is the kind that needs declaring, declare it. If it is not, there is nothing to hide. Building the decision around detectability rather than around whether you can defend the work is how authors end up submitting sentences they cannot explain when a reviewer pushes on them.
Not in the sense that prior publication means. That concept, as it is normally used, turns on whether the work was made publicly available in a citable, findable form, and sending text into a private tool is closer to emailing a draft to a colleague. The exception worth checking is any service that makes submissions visible to others, through a public feed, a shared workspace or a community gallery, because that genuinely does make the text public.
A manuscript usually reports aggregated results rather than identifiable records, which makes it a different decision from uploading raw data. Participant-level material, especially transcripts, open-ended responses and clinical notes, should be checked against your ethics approval and the promise in your consent form, not against general advice about AI. If you told participants their data would stay within the research team, that commitment has no convenience exception.
Treat this as a different question from using a tool on your own work, because it is. The manuscript belongs to authors who never consented to it leaving the confidential channel, and the editor invited you specifically for your judgement. The safe position is to keep the manuscript, and recognisable parts of it, out of external systems, and to ask the handling editor if you are unsure what the journal permits.
The most durable version I know is the difference between criticism and composition. A tool that tells you your sample size does not support your claim is doing what a colleague does in the margin, and the paper stays yours. A tool that produces the paragraphs is contributing content, which touches authorship and belongs in a declaration. Between them sits a gradient, so use a test rather than a rule: can you explain why every claim is true, and defend it without returning to the tool.