How to cite ChatGPT and other AI tools

Why AI citation is unusually messy

Every other source type has a settled convention because it has existed for decades. Generative AI has been in general student use since late 2022, and the style bodies published guidance at different times, from different assumptions, and have already revised it. So the honest answer to "what is the correct format" is that it depends on which style, which edition, and — more than for any other source — what your own department has decided.

Underneath the disagreement sits a genuine problem. A reference exists so a reader can retrieve what you read. Nobody can retrieve your ChatGPT session. The same prompt on the same model on the same day returns something different, and the model itself is replaced every few months. That is why the style bodies reached for the closest analogue they had — software, or personal communication — and why none of the results feel quite right.

The practical consequence is that a citation alone does not do the job. It tells your marker that a tool was involved; it does not let them see what the tool produced. That is why almost every institution now asks for the prompt and the output as well, usually in an appendix, and why the declaration matters more than the punctuation.

The format in each style

These are the current published positions of the style bodies. Where your handbook differs, your handbook wins — departmental guidance overrides the manual every time, and markers assess against the handbook.

Name the model, not just the company

The author is the organisation that made the tool — OpenAI, Anthropic, Google — and the title is the tool itself. Students routinely invert this and write "ChatGPT (2026)" as the author, which is wrong in every style: ChatGPT is the work, not its creator.

Include the version or the date of the specific model wherever the tool exposes one, because that is the only retrieval information that has any meaning. "ChatGPT" alone in 2026 could refer to half a dozen underlying models with materially different behaviour. Where no version string is published, the date of your session is the next best thing.

Do not cite an AI tool as a personal communication. APA considered and rejected that framing: personal communication is reserved for a human exchange, and treating a model as a correspondent implies an interlocutor who can be held to what they said.

Where the prompt and output go

Because your session is not retrievable, most institutions want the exchange itself preserved. Three conventions are common, and the difference between them is length.

• Short exchange — reproduce the prompt and the reply in the body of your text, set off as a block quotation with the citation after it.
• Longer exchange — put the full transcript in a numbered appendix and reference it in text: "(OpenAI, 2026; full transcript at Appendix B)".
• Many exchanges — a single appendix listing each prompt, the date, and a summary of what you took from the answer. This is the format most dissertation handbooks now specify.

Citing is not the same as being allowed

This is where students come unstuck. A perfectly formatted reference does not make a use permitted. Institutions in Australia, the UK, New Zealand, Canada and the US have converged on a three-tier model, and the tier is set per assessment, not per course.

The first tier prohibits generative AI entirely; the second permits it for defined support tasks with declaration; the third permits it broadly, again with declaration. Almost nowhere permits undeclared use, and almost nowhere permits submitting model-generated prose as your own writing regardless of whether you cite it.

Read the assessment brief rather than the course page. Where a brief is silent, ask the module leader in writing and keep the reply — a silent brief is not permission, and "the guidance was unclear" is not a defence that boards accept.

What a declaration usually asks for

Declarations vary in wording but almost all ask the same four questions. Answer them plainly; hedged declarations attract more scrutiny than honest ones.

• Which tool, and which version or date.
• What you used it for, in specific terms — "generated three possible structures for section 4", not "assisted with planning".
• What you did with the output — kept, rewrote, discarded, verified against sources.
• How you checked it. Naming the sources you used to verify a claim is the single strongest thing you can put in a declaration.

Never cite a reference the model gave you

Language models produce citations that look correct and do not exist. They combine a plausible author with a plausible journal and a plausible year, and the result passes every eye test except the one that matters, which is opening it.

The failure mode is specific and it is why markers can spot AI-assisted reference lists quickly: the DOI resolves to a different paper, the volume number and year do not match the journal's actual run, or the article title exists but under different authors. A marker who checks three references and finds one fabricated will check all of them.

So the rule is absolute. Every reference in your list must be one you have opened. If a model suggests a source, treat the suggestion as a search term, find the real record in your library database or on the publisher's site, and cite what you actually retrieved — with the details as the publisher gives them, not as the model gave them.

The uses that are almost always safe

Not everything is contested. Across the three-tier models, a consistent set of uses is permitted at tier two and above because none of them produce assessable content.

• Explaining a concept you then go and read about properly in the set literature.
• Generating practice questions to test yourself against.
• Checking your own sentence for grammar, without letting it rewrite your argument.
• Suggesting search terms and synonyms for a database search you then run yourself.
• Summarising your own draft back to you, as a check on whether your structure is legible.