Citing AI assistance honestly
Most AI-use policies leave you to write the disclosure yourself. Here is a defensible standard — and the exact phrasing for each common case.
he hardest part of an AI disclosure is not the principle. It is the wording. Most policies say something like "disclose any substantive use of AI tools" and then leave you to figure out what that sentence should actually look like at the bottom of your paper or the end of your essay. This piece gives you a defensible standard and the exact phrasing for the cases people get wrong.
The principle in one sentence
Disclose materially. If an AI tool shaped the substance — the argument, the evidence, the structure, or significant prose — say so, briefly, in the place your venue expects. The threshold is not "did I touch a model"; it is "would my reader's understanding of who did the thinking change if they knew". When in doubt, disclose. Under-disclosure is the cost that compounds; over-disclosure costs you nothing.
What to disclose, by case
Four common cases, with a phrasing for each:
- ❧Light editing only — the model fixed grammar or smoothed phrasing on prose you wrote: "The author used [Tool] for copyediting and grammar correction; all claims, structure and citations are the author's own."
- ❧Brainstorming or outlining help — the model suggested structures or framings you then chose between: "The author used [Tool] to explore possible structures during outlining. The final argument and prose are the author's own."
- ❧Substantive drafting assistance — the model produced prose you then edited heavily: "The author used [Tool] to draft initial passages of [sections]. Each passage was reviewed, revised and is defended as the author's own."
- ❧Translation from a first language — the model translated work you wrote in another language: "This paper was drafted in [language] and translated to English with [Tool]; the author has reviewed the translation."
Disclosure is not an admission of weakness. It is the move of a writer confident enough to show the working.
What you do not need to disclose
Reading a model summary of a paper you also read. Using a model as a thesaurus or to define a word. Asking a model to check whether a single sentence is grammatical. These are at the level of using a search engine or a dictionary, and no reasonable policy treats them as material assistance. The honest line is shaped by impact, not by paranoia.
Where the disclosure goes
Academic papers: in the methods section, in the acknowledgements, or in a dedicated AI-use statement when the journal asks for one — the ICMJE recommendations and Nature's editorial position both name the disclosure section as the place. Coursework: at the end of the document under a heading like "AI use", unless the assignment specifies otherwise. Cover letters and professional writing: almost never — but only because the convention is that you take direct responsibility for the words you sign, which means the threshold is whether you can defend each sentence, not whether you typed every word.
A short, honest sentence at the right moment protects you for the rest of your career. A buried use, or none at all when one was needed, is the kind of mistake that gets discovered later — by a journal, an HR review, or a colleague with a long memory. Disclose materially, and you sleep at night.
That was the last lesson in The Honest Editor's Desk.
Do I have to disclose every grammar fix?+
No. Light copyediting — the kind a friend would do — is below the threshold of material assistance for almost every policy. The line is whether the tool shaped substance, prose, or structure beyond the surface.
What if my institution has no policy?+
Default to disclosure when in doubt. A short, honest sentence costs you nothing and protects you against a policy that may be written, or applied, later.
Is naming the specific tool required?+
Not always, but it is the safer practice. "A large language model" is a defensible minimum; the specific tool name is better, because it lets a reader judge what kind of assistance you had.