CiteOwl

Blog/Using AI honestly

AI disclosure statements: what to actually write

The CiteOwl Team· Updated 20 September 2026· 18 min read

Write two or three sentences. Name the tool and version, the date, what you asked it to do, which part of the work it touched, and what you did with the output. Say whether you still have the records. A disclosure is not a citation, and no style guide governs it. Your institution does, so your assignment brief outranks any template online. Write the version you could read aloud in a meeting, then keep the drafts.

Citing AI and disclosing AI are two different jobs

The box on your submission page is not asking for a reference. A citation points at a specific output you are treating as a source, and takes whatever shape APA, MLA or Chicago gives it. A disclosure describes what you did while writing, and takes whatever shape your institution gives it. Two jobs, two authorities, and confusing them is most of why the box feels impossible to fill.

The MLA says this outright. In "Beyond Citation: Describing AI Use in Your Work", published 13 August 2025, it writes that "Acknowledgment of the use of AI is slightly different from, and more comprehensive than, citation", because "Citation points to an exact source of information" while "generative AI output is not a source in the traditional sense".

So the working rule is short. If you can point at a specific output, cite it. If you cannot, disclose it. Almost all real use produces nothing quotable, so most students need a statement and no reference. If you do need the reference too, the three styles disagree about it more than you would expect, and the current format for each is further down.

Find out which of five courses you are in

You cannot write the statement until you know which policy you are under, and that lives in your assignment brief. Policies land in roughly five places, and the collections universities publish for their own staff show how wide the range is.

Everything is banned. The University of Texas at Austin's collection of sample syllabus statements opens with a category whose bluntest line reads: "The use of generative AI tools for any part of your work will be treated as plagiarism." The University of Kentucky's collection of course policy examples carries one banning them "for any stage" of an assignment, brainstorming and grammar correction included.

Writing is banned, light editing is not. Johns Hopkins keeps a collection of syllabus statements written by its own instructors. One, for a course on the economics of sustainable energy, prohibits "using ChatGPT or other AI tools to generate written content for assignments" but allows them "for light editing purposes, focusing on improving grammar and clarity". Two verdicts in one statement, and only the second leaves you anything to declare.

Named tasks are allowed, and you must say so. Another Hopkins statement lets students "use generative AI to brainstorm ideas or ask questions about how to solve a problem", adding that anyone who does "should indicate how they used it in their homework or paper". Kentucky's conditional category is more specific: allowed "for planning, organizing, outlining, brainstorming, and copy-editing", not for "composition, writing, or content generation", and the final work must be "at least 75% your work".

Allowed, but show the whole exchange. Both the Texas and the Northern Illinois collections carry a policy asking for "an appendix showing (a) the entire exchange, (b) a description of precisely which AI tools were used, (c) an explanation of how the AI tools were used, (d) an account of why AI tools were used". Texas's most permissive category asks instead for "a paragraph that explains what AI content-generation tool you used, the dates you used it, and the prompts you used", and wants AI passages set "in a different colored font".

Nothing is required at all. The Northern Illinois collection also reproduces a University of Delaware option telling students that "no special documentation or citation is required". Rarer than the internet suggests, and still a policy: the course decided, and you are following it.

Which means any template you find online is answering somebody else's course. Read the brief first, then pick the sentence.

The rule in your brief is probably new, and the numbers explain why. HEPI's Student Generative AI Survey 2026, published 12 March 2026 from 1,054 full-time UK undergraduates surveyed in December 2025, found 94% using generative AI to help with assessed work, up from 89% a year earlier and from just over half two years before, and the share putting AI-generated text straight into assessed work rising from 3% in 2024 to 8% in 2025 to 12%. A policy written against those numbers is a decision, not boilerplate, so read it as one.

The six things a disclosure has to say

Once you know you owe a statement, the contents are the least controversial part of the subject. Institutions on three continents, and the journal publishers, ask for the same fields.

• Which tool, and which version. Sydney asks for the name and version, the publisher and the URL. Taylor & Francis wants "the full name of the tool used (with version number)".
• When you used it. The University of Queensland's acknowledgement guide lists the date as a required element.
• What you asked it to do. Queensland wants the "prompt(s) used". Cambridge's School of the Humanities and Social Sciences publishes a template declaration whose third question is "What prompt did you provide?"
• Which part of the work it touched. Queensland asks for the "part, section or page of the assessment piece".
• What you did with the output. Cambridge's fifth question: "How have you used or changed the generative AI's output?"
• Whether you still have the records. The University of Melbourne asks you to state "whether detailed records of prompts and outputs you used in the AI tools or technologies are available on request".

Assembled, that is one reusable sentence:

I acknowledge using [tool, model or version, link] on [date or date range] to [what you asked it to do]. I used the output to [what you did with it], in [which part of the work]. [What in the submitted work is yours, and what is not.]

Melbourne's own examples close with "A full record of prompts and outputs is available upon request", and several institutions expect it. Add it only if it is true: it is a promise about files on your machine, and it becomes a false statement the moment you clear a chat you said you had kept.

The statements themselves

Five cases, in the order students meet them. Each is short enough to paste, and each carries the thing no other template page tells you: which policies would still refuse it. A disclosure describes your use. It does not authorise it.

You did not use AI at all

I did not use any generative AI tools in preparing this assignment.

Melbourne's version is "I have not used any AI tools or technologies to prepare this assessment" and RMIT's is "I did not use any AI tools in the creation of this work". Many courses ask every student for a statement either way, and an empty box is ambiguous where a sentence is not.

Read the scope first. The University of Sydney requires you to acknowledge "any tools that use generative AI, such as translation tools, paraphrasing tools or referencing tools", exempting only tools "used for word processing, or which only correct basic spelling and grammar". A paraphraser, a translator and a bibliography generator all count, so this template may not be true for you however much it feels like it is.

Would a strict policy refuse it? No. This is the one statement no course can object to, which is exactly why it has to be accurate.

You used AI for brainstorming or structure only

I acknowledge using ChatGPT (GPT-5, chatgpt.com) on 14 March 2026 to brainstorm angles on congestion charging and to question the order of my three sections. I used the output to decide what to argue and in what order. No text from the tool appears in this assignment. The wording, the sources and the argument are mine.

The last two sentences are what make it useful. A marker reading "I used AI for brainstorming" cannot tell whether that was a five-minute conversation or a machine-written outline you filled in. Saying which costs a clause.

Would a strict policy refuse it? Yes, under a total ban, and specifically under the Kentucky example that names brainstorming as prohibited. It is exactly what is permitted, and exactly what must be declared, under the Hopkins statistics statement above.

You used AI to edit your own prose

I acknowledge using Claude (Sonnet 4.5, claude.ai) on 2 April 2026 to edit prose I had already written in sections 2 and 3, for grammar, sentence length and clarity. I gave it my own draft and asked it to flag unclear sentences. I accepted some suggestions and rejected others, and every sentence submitted is one I approved. No content, sources or arguments came from the tool.

Be careful with the word editing, because it covers two things. A spellchecker and a language model rewriting your paragraph are not the same object. Publishers draw the line explicitly: Elsevier says "basic checks of grammar, spelling and punctuation do not need a declaration statement", and Springer Nature exempts "AI assisted copy editing". Both were written for manuscripts, and a model that reshapes a sentence has gone past them. If you asked it to make the writing better rather than to fix errors, write that instead.

Would a strict policy refuse it? Yes, under the Kentucky no-use example, which names grammar correction. It is expressly permitted under the Hopkins sustainable energy statement, which allows light editing for grammar and clarity and nothing else.

You used AI to find and summarise sources

I acknowledge using Elicit and ChatGPT (GPT-5) between 3 and 9 May 2026 to find and summarise research on remote work and productivity. I used them to locate candidate papers and to read short summaries before deciding what to open. I then read every source I cite and checked each claim against the paper itself. Two suggested references did not exist and were discarded. No summary text from the tools appears in the assignment.

Every sentence here is a factual claim about your own behaviour, and the third is checkable in ten minutes by anyone who opens a reference and asks you about it. Do not write it unless you did it, and if you have not, check the citations first.

Would a strict policy refuse it? Yes, under any total ban. And when the searching is part of your method, as in a literature review, the tool belongs in your methodology section too.

AI drafted a passage that you then rewrote

I acknowledge using ChatGPT (GPT-5, chatgpt.com) on 18 May 2026 to draft an initial version of the section on grid storage costs, pages 4 to 5. I prompted it with my own outline and notes. I then rewrote the passage in my own words, corrected two factual errors, and replaced its general claims with sources I read myself. The structure and the first draft of that section originated with the tool.

This is the case most students have and the one most policies refuse. Kentucky's conditional example excludes "composition, writing, or content generation", and one of its no-use rationales counts even "rewriting generated outputs completely in your own words" as misuse.

Would a strict policy refuse it? Almost every policy except an explicitly unrestricted one, or a course that asks for the full transcript anyway. Under anything else this is not a permission slip. It is a clean, dated, signed account of a breach. The fix is to write the section yourself before you submit, never to soften the sentence, because softening it turns a fixable problem into a false declaration.

Where it goes, and whether it counts

Institutions disagree about placement more than about contents, so the brief wins here too. Here is the range, and a default for when nothing is specified.

Some courses skip all of this and give you a form: one Hopkins statement simply tells students to "disclose via this form", so the box you are looking at may not be in your document at all.

If nothing is specified, use Melbourne's default. Its guidance says a declaration "can be added at the end of your assessment", "generally follows the reference list and does not need to be started on a new page", and "is generally not included in your word count". That last clause answers the question everyone actually has, which is whether being honest costs you words you needed.

If you are submitting to a journal

Publishers want the same six facts and put them somewhere else, which is why a journal-shaped declaration pasted into a course assignment reads wrong, and vice versa.

Elsevier wants "a separate declaration section at the end of your manuscript, immediately before the references", opening "During the preparation of this work the author(s) used [NAME OF TOOL/SERVICE] in order to [REASON]" and closing with a confirmation that you "reviewed and edited the content as needed and take(s) full responsibility for the content of the publication". Springer Nature puts it in the Methods section instead, and Taylor & Francis with the article.

All three agree on the thing students sometimes get wrong. No AI can be listed as an author. Elsevier's reasoning is the clearest: "Authorship implies responsibilities and tasks that can only be attributed to and performed by humans." Your name on the work is the accountability, which is also why your course wants a statement rather than a footnote.

If you do owe a citation as well: APA, MLA and Chicago

A disclosure describes what you did. A citation points at a specific output you are treating as a source, and only the second one has a style manual behind it. Editions first, because half the guides online format against a superseded book. APA is on the Publication Manual, 7th edition (2020), MLA on the MLA Handbook, 9th edition (2021), and Chicago on the 18th edition (2024), the first Chicago edition to cover AI in the book itself, at section 14.112. All three have published AI guidance since their manual shipped.

They do not agree, and the disagreements are the part no single style site shows you.

The second row is the one that catches people. A student who learned "cite ChatGPT like any other source" from an APA guide, then submits to a Chicago-styled history course, writes the bibliography entry Chicago specifically tells you not to write.

APA

The current template comes from "Citing generative AI in APA Style: Part 1", published 9 September 2025: AI Company Name. (year, month day). Title of chat in italics [Description, such as Generative AI chat]. Tool Name/Model. URL of the chat.

Anthropic. (2026, June 20). Urban heat islands and tree canopy cover [Generative AI chat]. Claude Sonnet 4. https://claude.ai/share/…

In text it behaves like any author-date source, (Anthropic, 2026) or Anthropic (2026). If a guide hands you OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model], it is showing you a retired shape: APA updated that 2023 post on 15 September 2025 with a notice that it "is retained for archival purposes only".

MLA

MLA assembles the entry from its usual core elements. Per the MLA Style Center's revised guidance of August 2025, describe what was generated in the title slot, usually by quoting your prompt; the tool is the container; the Version element carries the model as specifically as possible; the company is the publisher; and the location is the stable, shareable URL.

"Summarise the evidence on cycling infrastructure and commuter mode share" prompt. ChatGPT, model GPT-4o, OpenAI, 7 Apr. 2025, chatgpt.com/share/…

The in-text citation points at the first words of the entry, so it reads ("Summarise the evidence").

Chicago

Chicago treats generated text closer to a conversation than a publication. Its own Q&A says not to put a chatbot in a bibliography or reference list unless you can provide a publicly available link, because otherwise the exchange is no more retrievable than a phone call. The credit goes in your text or a note.

1. Text generated by ChatGPT, OpenAI, June 20, 2026, https://chatgpt.com/share/…

For author-date papers the parenthetical is (ChatGPT, June 20, 2026). And the rule people skip: if you edited the generated text, say so, "edited for style and content" at the end of the note, because an unmarked edit quietly changes who is responsible for the sentence.

Where your course has not specified, our own judgement is to disclose more than the manual asks and cite less. A chatbot entry is the weakest line in any bibliography, sitting in the same alphabetical list as papers that went through review while pointing at something nobody else has checked and nothing can re-run. The exception is narrow and real: when the output itself is what you are discussing, quoting or analysing, it is evidence, and then it belongs in the list with a date and a share link.

The statement is yours, and that is the whole risk

Everything above treats the disclosure as an administrative field. It is not. It is a signed claim about your own conduct, filed with the work it describes, and it outlasts the mark.

The consequences are written down where you would expect. Loughborough tells students that if you are "suspected of inappropriately using AI or have failed to acknowledge its use, this will be regarded as academic misconduct by the University and you will be referred to one of its Academic Misconduct Committees". Cambridge says a student "using any unacknowledged content generated by artificial intelligence within a summative assessment as though it is their own work constitutes academic misconduct".

Notice what both punish. Not the use, the silence about it. If your course permits the tools, you are not building a defence, you are describing five minutes of your afternoon.

A disclosure also only helps when the underlying use was allowed. Declaring a banned use documents it cleanly, it does not license it: confirm the tool was permitted, then say how you used it. We work through where courses draw that line in is using AI to write essays cheating.

Which makes the real work a habit rather than a sentence. If you know, at the end of a paper, which passages you drafted, which you revised with help, and which sources you opened yourself, the statement writes itself in ninety seconds. If you do not know, no template can rescue you, because the missing thing is not the wording.