AI Disclosure and Citation: What Students Actually Need to Do

Whether using AI on an assignment counts as misconduct is not a philosophical question. It is a documentary one. There is a specific piece of writing that governs your specific submission, and it either permits AI use, restricts it to named tasks, or forbids it. Most students never read that document. They ask a friend, they read a news article about a university somewhere else, or they assume the rules match whatever their last module allowed.

The citation side is the opposite problem. Students agonise over how to cite AI as if the answer were unsettled, when APA and MLA have both published formats and the mechanics are fixed. This article separates the two: the part that varies and that you must check, and the part that is already decided and that you can just copy.

The policy hierarchy, and which document actually wins

Rules on AI use are usually written at several levels at once, and they do not always agree. Understanding the order matters because a general permission at one level does not survive a specific prohibition at another.

1. National or sector-level guidance. Some countries have qualification authorities or funding bodies that publish expectations for AI use in assessed work. Usually broad and advisory rather than enforceable against an individual student.
2. Institutional policy. Your university academic-integrity policy. This sets the outer boundary and defines the penalties, then usually allows departments to vary the default.
3. Departmental or programme rules. A law school and a computer science department in the same university frequently take different positions, and both are legitimate. Programme handbooks are where this appears.
4. Module or course handbook. Narrower again, and typically written by the person teaching it.
5. The assignment brief. The most specific document, and in practice the one that wins.

The most specific instruction governs. If your institutional policy permits AI with disclosure but the brief for this essay says no AI tools of any kind, the brief applies. That is not a contradiction: the institution set a default, and the assessor exercised the discretion the policy gave them, usually because the assignment tests exactly the skill the tool would replace.

The reverse also happens, and students get it wrong more often. A brief that says nothing about AI does not grant permission. Silence at the brief level means you fall back to the module handbook, then the programme rules, then institutional policy. Silence is not consent; it is a pointer upward.

The three common policy positions

Almost every policy you will encounter is a version of one of three positions. Identifying which one you are under tells you what you need to do.

1. Prohibited

No generative AI at any stage. This is most common for language assessments, timed work, foundational skills modules, and anything where the writing itself is the thing being examined. Under a prohibition, using a tool to "just fix the grammar" is still a breach. So is using it to generate an outline you then wrote from. The scope is usually the whole process, not just the final text.

2. Permitted with disclosure

You may use AI, and you must say so. This is now the most common position at institutions that have settled their policy. The disclosure requirement is typically satisfied by a short statement, sometimes in a specific place (a cover sheet, an appendix, a designated box in the submission system). Two things go wrong here: students disclose vaguely ("AI was used"), and students disclose in the wrong place, which some systems treat as non-disclosure.

3. Permitted for specified tasks only

The most detailed version, and increasingly common. The policy lists what is allowed and what is not. A typical split allows brainstorming, generating practice questions, explaining a concept you did not understand, and grammar or spelling correction — while prohibiting drafting prose that appears in the submission, generating analysis or arguments, summarising sources you have not read, and producing citations or reference lists.

The line in this third category is usually between support for your thinking and substitution for your thinking. If the tool produced a sentence, an idea, or a conclusion that ended up in your work, that is normally the prohibited side, even when a proofreading permission exists.

If the policy genuinely is silent

Sometimes you check all five levels and find nothing. In that case, ask — and ask in a way that produces a useful answer rather than a defensive one.

Bad question: "Can I use ChatGPT?" This is unanswerable. It invites a cautious no, or a vague yes that protects nobody.

Good question names the task, the tool, and the stage:

For the 2,500-word case analysis due on the 14th, I would like to use ChatGPT for two things: generating a list of possible counter-arguments after I have written my own analysis, so I can stress-test it, and checking grammar in my final draft. I would not use it to draft any of the submitted text or to find sources. Is that acceptable, and if so would you like me to include a disclosure statement? Happy to follow whatever format you prefer.

Three things make this work: it is specific about the tasks, it rules out the uses instructors worry about, and it offers to disclose. Send it by email rather than asking after a lecture, so the answer is in writing, and keep the reply. A dated email from the module leader is worth far more than your recollection of a conversation.

Citation formats

If your use requires citation rather than only disclosure — for example, you quoted or drew on model output as a source — both major style guides have published guidance.

APA 7

APA treats a large language model as software. The reference-list entry credits the organisation that made it, not the model itself as an author:

OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model]. https://chat.openai.com/chat

In-text, that becomes (OpenAI, 2023) or "OpenAI's (2023) ChatGPT…". Use the year of the version you actually used, and name the version in parentheses as shown.

APA also makes a point that matters more than the punctuation: model output is not retrievable by your reader. If someone follows the URL and enters your prompt, they will not get your response back. A citation normally lets a reader verify a source, and this one cannot. APA's recommendation is therefore to include the prompt you used and the response you received — either in the body of the text, if it is short, or in an appendix if it runs long. Cite the appendix from the relevant point in your text.

MLA 9

MLA works from its template of core elements rather than a fixed pattern, and its guidance is to describe the AI use rather than treat the model as an author. A works-cited entry describes the prompt in the title position, then names the tool, its version, the maker, the date, and the URL:

"Explain the difference between formative and summative assessment" prompt. ChatGPT, 14 Mar. version, OpenAI, 22 Feb. 2024, chat.openai.com/chat.

MLA also advises citing AI use in your text where relevant and vouching for any information you take from it — which is a polite way of saying that if the model states a fact, verifying it is your responsibility, not the model's.

Comparing the two

Whichever style you use, check your institution's own guide first. Some universities publish a house variation, and where they do, theirs governs.

Sample disclosure statements

These are templates to adapt, not to paste unchanged. Specificity is the point — a vague disclosure often reads worse than none, because it suggests you are minimising something.

(a) Brainstorming only

I used ChatGPT (OpenAI, GPT-4, accessed 3 March 2025) at the planning stage of this assignment. I asked it to suggest possible angles on the research question and to list common objections to the position I was considering. I used two of its suggestions as starting points for my own reading. No text generated by the tool appears in this submission, and all sources cited were located and read by me.

(b) Grammar and language editing only

After completing my final draft, I used ChatGPT (OpenAI, accessed 3 March 2025) to check grammar, punctuation and sentence clarity. I reviewed each suggested change individually and accepted approximately half of them. The structure, argument, evidence and wording of the substantive content are my own. No content was generated by the tool.

(c) Drafting assistance, then substantially revised

I used ChatGPT (OpenAI, GPT-4, accessed 3 March 2025) to produce an initial draft of the background section (pp. 2–4) from an outline and set of notes I wrote. I then substantially rewrote that draft: I verified every factual claim against the cited sources, removed two claims I could not verify, restructured the section, and rewrote the prose in my own words. The remaining sections were written without AI assistance. The prompts used and the original generated text are included in Appendix B.

Note what statement (c) does. It names the sections affected, describes the revision honestly, admits that some generated claims did not survive verification, and provides the raw material for scrutiny. That is a disclosure an assessor can work with. "I used AI to help with some of the writing" is not.

Keep your drafts, always

This applies regardless of whether you used AI at all, and it is the single most useful habit in this whole area. Version history is the only thing that reliably demonstrates authorship.

• Write in an editor that keeps revision history automatically. Google Docs, Word with OneDrive or SharePoint, Overleaf and Notion all do; local-only files with no backup do not.
• Keep your notes, annotated PDFs and outlines. A chain of working materials is more persuasive than any single document.
• If you used AI, save the transcripts including prompts as PDFs — chat histories can be deleted or lost when accounts change.
• Keep the email in which your instructor answered your question about permitted use.

Students who get into difficulty are rarely undone by the AI question itself. They are undone by having no record of how the work came to exist. Detector scores are difficult to argue with in the abstract; a document that grew across fourteen sessions over three weeks is not. If you are curious about why those scores are so contestable, it is worth understanding what bypassing AI detection actually involves and why the whole framing tends to mislead.

The short version

Find the assignment brief. If it addresses AI, follow it. If it does not, work upward through the module handbook, programme rules and institutional policy until you find a position, and follow that. If nothing addresses it, email your instructor a specific question and keep the reply.

Then, if you used AI in a way that requires acknowledgement, cite it in your required style and write a disclosure statement that says what you actually did. The formats are published and stable. The judgement call is not the citation — it is being accurate about your own process, and keeping the evidence that shows you were.