Generative AI arrived in scientific writing faster than the guidance did, and the guidance that exists is spread across publisher policies, journal instructions and the occasional editorial. The result is a lot of researchers who are unsure whether they are allowed to use these tools at all, and a smaller number using them in ways that will cause them a problem.
Both positions are avoidable. The settled parts of this are more settled than they look, and the risks are specific rather than general.
Two cautions about this page. Journal policies on AI are being revised more often than almost anything else in publishing, so check the policy of the journal you are submitting to rather than relying on any summary, including this one. And the date at the bottom of this page is there for a reason.
The parts that are settled
An AI tool cannot be an author. This is the clearest rule in the area and it follows from what authorship means rather than from any view about the technology. Authorship requires accountability for the work: being able to vouch for its accuracy and integrity, and to approve the final version. A software tool cannot be held accountable for anything, so it cannot qualify, no matter how much of the drafting it did. The International Committee of Medical Journal Editors sets out the authorship criteria this rests on, and the major publishers have all converged on the same conclusion.
You remain fully responsible for everything in the paper. Including every sentence a tool produced and every reference it supplied. “The model wrote it” is not a defence that exists.
Use has to be disclosed. Nearly every significant publisher now requires authors to say whether and how AI-assisted tools were used. The common pattern is that help with writing, editing or language goes in the acknowledgements, while use of AI to collect or analyse data, or to generate or process images, goes in the methods, with the tool and version named. Some journals also want it in the cover letter. The distinction behind this is worth internalising: assistance with expressing your work is a disclosure, whereas AI in the generation or analysis of your results is part of the methodology, and that is a difference in kind.
Generated images are treated much more strictly than generated text. Several major journals prohibit AI-generated or AI-altered images in figures outright. Enhancement of a micrograph by a generative model is not a presentation choice; it creates data that was not measured.
Where the real risks are
Fabricated references. This is the failure that has embarrassed the most people. Language models produce citations that are correctly formatted, plausibly titled, attributed to real authors who work in the right field, and entirely non-existent. They are difficult to spot precisely because they look right, and a reviewer in the field will spot them immediately. Check every reference a tool gives you against the actual paper. Not the title: the paper.
The same applies to any factual claim. A confident, fluent, specific sentence about what a 2019 study found is exactly the output these systems produce most readily when they do not know.
Confidentiality. Pasting an unpublished manuscript into a general-purpose AI service means disclosing it to a third party, and depending on the terms, possibly contributing it to training data. For your own manuscript that is a decision you are entitled to make, though your co-authors and your institution may have views, and a collaboration agreement or a clinical data consent may well forbid it.
For a manuscript you are reviewing it is not your decision. A manuscript under review is confidential, you received it on that basis, and uploading it to an external tool is a breach of that confidence regardless of outcome. Publisher policies are increasingly explicit about this, and some prohibit AI use in peer review altogether. It is worth being careful about this even where no policy has caught up, because the obligation does not come from the policy.
This is also why I do not put client material into AI writing tools. Someone sending me unpublished work has to be able to know where it goes.
Flattening. The quieter cost, and the one I would raise first with a researcher. These tools produce prose that is fluent, even, and generic. It smooths out the places where a paper is making a precise claim, and it has a strong pull towards hedged, padded, conventional phrasing. A paper written this way reads as competent and says less than it meant to. The specific claim, stated plainly, is the thing that makes an editor sit up, and it is the first thing a general-purpose model will soften.
If you use a model to improve a draft, read the result against what you were actually trying to say, and put back the precision it removed.
Uses that are straightforwardly fine
Language help, particularly if you are not writing in your first language. This is the strongest case for these tools and it is close to an equity argument: a researcher whose English is their third language has always been disadvantaged in a way unrelated to the quality of their science, and a tool that narrows that gap is doing something useful. Disclose it as you would any language editing, and read the output carefully, because models do sometimes change meaning while improving grammar.
Asking what is unclear. Giving a model your abstract and asking it what the central claim appears to be is genuinely informative, because if it comes back with the wrong answer, a hurried reader may too.
Mechanical work: reformatting references, checking consistency of terminology, converting a document to a journal’s structure, drafting a plain-language summary that you then rewrite.
On AI detectors
If you are accused of using AI because a detector said so, the evidence is weak, and it is worth knowing that.
These tools are unreliable in both directions and they are documented to be biased against text written by non-native English speakers, whose writing tends to have exactly the lower lexical variation the detectors read as machine-generated. They also flag heavily edited prose and the clean, conventional register that scientific writing rewards. There is no reliable way to determine from the text alone whether a passage was machine-written.
If you are on the receiving end of this, ask what the evidence is, and say plainly what you did and did not use. If you are in a position to make such an accusation, a detector score is not sufficient grounds.
A practical position
Decide what you will use these tools for before you are tired and close to a deadline, which is when the bad decisions get made. Keep a note of what you used and for what, so the disclosure is a two-minute job rather than a reconstruction. Check the journal’s policy at the point of submission rather than relying on what it said last year. Never put a manuscript you are reviewing into one. And verify every reference, every time.
Used that way they are a drafting aid, which is a useful thing to have. The failures all come from treating them as a source of facts or as a substitute for the thinking.