How to Cite and Disclose AI Use in Academic Work

Disclosure is the cheapest protection you have

Most academic misconduct cases involving AI are not really about the AI. They are about a student who used it, said nothing, and then had to argue about it afterwards. Declared use is almost never the problem; undeclared use is.

Before anything else: read your institution''s policy. Rules vary enormously — some departments permit AI for brainstorming and editing but not drafting, others ban it entirely for assessed work, others require a signed statement. The style guides below tell you how to cite; only your institution tells you whether you may.

The three major formats

APA (7th edition)

APA treats generative AI as software, citing the company as author and the model as the title.

Reference list: OpenAI. (2026). ChatGPT (version) [Large language model]. https://chat.openai.com

In text: (OpenAI, 2026)

APA also asks you to describe how you used it in your method or introduction, and recommends including the prompt and response in an appendix where the output itself is being discussed.

MLA (9th edition)

MLA templates it around the prompt as the source title.

Works cited: "Describe the causes of the 1929 crash" prompt. ChatGPT, version, OpenAI, 12 Mar. 2026, chat.openai.com.

In text: ("Describe the causes")

MLA explicitly advises that you cannot treat the tool as an author and that you must verify any facts it produced against real sources.

Chicago (17th/18th)

Chicago prefers a note rather than a bibliography entry, since the content is not retrievable by readers.

Footnote: Text generated by ChatGPT, OpenAI, March 12, 2026, https://chat.openai.com.

The rule that matters more than the format

An AI model is not a source. It is a tool that produces text, sometimes accurate and sometimes confidently fabricated. Citing ChatGPT for a fact is roughly like citing "a conversation I had with someone at a party" — it documents where you heard it, not whether it is true.

So: if the model gave you a claim, statistic, or reference, find the primary source and cite that instead. Models fabricate citations fluently, including plausible authors, real journals, and DOIs that resolve to nothing. Checking every reference is not optional.

Writing a disclosure statement

A good statement is short, specific, and states what you did yourself. For example:

"I used ChatGPT (OpenAI, 2026) to generate an initial outline and to suggest counterarguments in section 3. All sources were located and verified independently, and all analysis and final text are my own."

Three elements make it work: the tool named, the specific task it performed, and an explicit statement of what remained yours. Vagueness here is what invites questions.

What disclosure does not do

• It does not make prohibited use permitted. If your institution bans AI-drafted submissions, disclosing that you drafted with AI is a confession, not a defence.
• It does not transfer responsibility. You remain accountable for every fabricated citation and every wrong claim in your submission.
• It does not exempt you from detection reviews. Though it makes them far easier to resolve — which is precisely the point.

Protecting yourself alongside disclosure

• Keep version history enabled. A document showing gradual development is the strongest evidence of authorship there is, as we detail in the false-accusation guide.
• Save your prompts and outputs. If you are asked to demonstrate what the tool actually contributed, having the transcript settles it in minutes.
• Verify every reference by opening it. Not by checking that it looks plausible.
• Know that detection is imperfect in both directions. Honest work gets flagged, particularly for second-language writers — understanding how these tools work helps you discuss the result calmly if it happens.

The short version

Check your policy. Use AI only where permitted. Cite it in your required style. Verify everything it tells you against a real source. Say plainly what you did. Keep your drafts. Do that and the question of what a detector thinks becomes almost irrelevant — you have the one thing no score can override, which is a documented record of having done the work.