Write an AI Disclosure Statement in 4 Lines, With 3 Worked Examples

4 lines, and each one answers a question a marker would otherwise have to ask. The tool, named with its version. What it was asked to do. Where its output sits in the submitted text, by section or paragraph. What was done to that text afterwards. A fifth line is optional and worth adding wherever a record exists: the chat link, or the document history.

A sentence version exists at university level.

The writing center at the University of Tennessee, Knoxville publishes a 2-sentence acknowledgment that names the tool, the company that made it, and how it was used in the assignment ¹. It is a formal sentence pair rather than a table, and it asks for the same facts.

The Task and After lines are the ones students leave out.

MLA guidance asks a writer to acknowledge functional uses of a tool, such as editing prose or translating words, even where nothing it generated is quoted ². Chicago tells a writer who reworked generated text to say so, and gives "edited for style and content" as an example of the wording ³.

Both of those are edits. Both still count.

About 60 words is the ceiling that works, a target from this article rather than from any style guide.

A statement that takes 10 minutes to write gets skipped. A skipped statement is worth nothing to the person reading the paper. Where the output also has to appear in the reference list, that is a separate job, worked through in citing ChatGPT in APA and MLA.

The syllabus decides how much the statement has to carry, and most course policies fall into 4 shapes. Always allowed, allowed on a stated condition such as attribution, banned except for named uses, or banned outright. A silent syllabus is the fifth case. The 4 lines stay the same in all of them. What changes is how specific the Task and Where lines have to be, and whether a statement is required at all.

The syllabus is the document a marker reads first, and where a campus-wide policy exists the syllabus sits under it.

The writing center at the University of North Carolina at Chapel Hill tells students to consult each syllabus for the requirements on AI tools ⁴. In the 2026 Time for Class survey, 32% of US higher education administrators reported an institution-wide AI policy ⁵, so at many institutions the course rule is the one a student can find, and 4 courses in 1 term can hand a student 4 different answers about the same grammar checker.

A condition attached to permission is the part that decides the statement.

"Allowed with attribution" means the Task line has to match what was asked for, in enough detail that a marker can see the permission covered it. Which of the 4 shapes a given syllabus states, and what silence in a syllabus means, is worked through in what an AI policy in a syllabus means.

3 uses cover most of what students do, and each one produces a different statement. Asking a tool for ideas and then writing everything alone. Running a finished draft through a grammar checker. Letting a tool rephrase sections that were already written. The 3 examples below are written out at the length a real one should be.

Ideas only. ChatGPT (GPT-5) was asked for 6 possible angles on the prompt before drafting started. None of its wording appears in this essay. The argument in section 2 began from its third suggestion and was built from the course readings. Chat link available on request.

Grammar pass. Grammarly was run over the finished draft for spelling, punctuation and comma placement. No sentence was rewritten by the tool, and no generative feature was switched on. Every suggestion was accepted or rejected one at a time.

The line about generative features is the one doing the work.

Grammarly states that, on its own AI detector, ordinary red and blue underline corrections should usually not move the score, while using its generative features to rewrite whole sentences and paragraphs will raise it ⁶, and that a score on one detector says nothing about the score on another. A statement that says grammar pass without saying which of the 2 was used has not said much.

Rephrased sections. ChatGPT (GPT-5) was given paragraphs 3 and 4 of a finished draft and asked to make them clearer. The returned wording was used as a starting point and then rewritten. Roughly half the phrasing in those 2 paragraphs came back from the tool. The original paragraphs are in the document history.

The third case is the one that decides a grade under a strict policy, and it is the one a vague statement hides.

12% of UK undergraduates surveyed in December 2025 said they put AI-generated text directly into assessed work, up from 8% a year earlier and 3% in 2024 ⁷.

Naming the paragraphs turns that from something found into something declared. Whether a grammar pass on its own breaks a course rule is a separate question, taken up in is a ChatGPT grammar check cheating.

At the top of the submitted file, above the first line of the essay, or in the submission comment box where the platform has one. Written during the work rather than after it. A statement composed at submission is a memory exercise, and memory can't be checked later against a chat session nobody saved.

APA notes that ChatGPT generates a unique response in each chat session even when given the same prompt ⁸, so re-running the prompt later does not reproduce the text that was used; only the saved session, or a copy of it, does.

Save the link, or paste the exchange into a document the same day it happens.

A running log is a university-level recommendation rather than one strict instructor's demand. The writing center at the University of Tennessee, Knoxville recommends that students in courses allowing AI keep a log of that use, and notes that the university has recommended instructors ask for the documentation ¹.

The log is also what makes the Where line writable at all. A student who noted "ChatGPT, thesis angles, 14 September" while working already has 3 of the 4 lines. Nothing noted, and the job becomes reconstructing 3 weeks of work from a browser history.

A draft written in Google Docs, or in Word, keeps a record of a different kind, and it comes with conditions. Microsoft states that version history in Microsoft 365 works only for files stored in OneDrive or SharePoint ¹⁰, so a Word file kept only on a local drive has no Microsoft 365 version history to show. Google states that browsing earlier versions of a file needs edit permission on it, so a document shared as view-only has not shared its history ¹¹. What that history shows, and the several ways it shows nothing, is set out in Google Docs version history as proof.

It cannot prove itself. A student set on hiding a use writes a false line as easily as a true one, and nothing in the form catches that. What the statement changes is the default. Honest work gets somewhere to be recorded, and a concealed use now needs a specific written claim that the paper can be read against afterwards.

It is also not a shield against a detector reading, and treating it as one sets a student up for a bad afternoon.

A flag describes text. It doesn't describe a person, and no reading settles who wrote a paragraph.

Detector numbers sit in that same frame, including the good ones. The tool at human.olive.is, called Human, reports an in-house measurement of its own, on a held-out set of its own and with no independent audit behind it: 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%, measured as of September 2026. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A ceiling drawn across 600 papers still settles nothing about which sentences in 1 of them a student wrote.

That fear is measured.

Among UK undergraduates surveyed in December 2024, the main things putting students off using AI were being accused of cheating, named by 53% of respondents, and false results, at 51% ⁹.

A statement does not remove that fear, but it gives a student somewhere to stand: a record to point at instead of a score to argue about, which is the part of the afternoon a student controls.

What happens when a flag lands anyway, and which evidence answers it, is in how to prove an essay was not written by AI.