How to cite ChatGPT and other AI tools

Before any of the formatting below: check whether you are allowed to use it at all.

Institutional policies on generative AI vary from “cite it like any other source” to “any use is academic misconduct”, and they vary by department within the same university. A correctly formatted citation of a tool you were not permitted to use is still a disciplinary matter. Your module handbook governs; nothing on this page overrides it.

With that settled, the two dominant styles have taken genuinely different positions, and the difference is more interesting than it first appears.

APA 7: the model has an author

APA treats AI output as software output, so the company that made the model is the author.

The pieces:

• Author — the organisation that made the tool, not the tool’s name. Anthropic for Claude, Google for Gemini, OpenAI for ChatGPT.
• Date — the year of the version you used.
• Title — the tool name, italicised.
• Version — in parentheses. OpenAI labels versions by date, so APA’s examples use the date form.
• Descriptor — [Large language model] in square brackets.
• URL — where the tool lives.

Why it is not a personal communication

The obvious objection is that nobody else can retrieve your particular conversation, and APA normally handles unretrievable sources as personal communications — cited in text only, with no reference entry.

APA explicitly rejected that reasoning here. Their position is that with AI-generated text “there is no person communicating”, so quoting it is more like reporting the output of an algorithm than reporting a conversation. An algorithm has an author, and authors get reference entries.

APA also recommends that because the output is not retrievable, you should include the full prompt and response in an appendix or as supplementary material, so a reader can see what you actually got. In practice this is the part students skip and markers notice.

MLA 9: the prompt is the title

MLA reached the opposite conclusion. In MLA there is no author at all — the tool is treated as a container, and the thing being cited is your prompt.

The pieces:

• Title of source — a description of your prompt, in quotation marks, followed by the word prompt.
• Title of container — the tool, italicised.
• Version — the model name or number. MLA specifically wants the model, so “GPT-4o version” rather than a date.
• Publisher — the company.
• Date — when you generated it.
• Location — a stable shareable link to the conversation if the tool offers one, otherwise a general link to the tool.

That preference for a shareable conversation link is worth acting on. Several tools now produce one, and it turns an unretrievable source into a retrievable one, which resolves most of the objection to citing AI in the first place.

Harvard: ask, then follow the pattern

There is no single Harvard authority — it is a family of styles, and your institution’s guide is the specification. Most university Harvard guides have now published AI-specific guidance, and where they have not, the convention that fits Harvard’s existing logic is:

If your library has published something different, use theirs. This is one area where “my citation generator said so” carries no weight at all.

Citing is not the same as acknowledging

This is the distinction that causes the most trouble, and MLA has been explicit about it: acknowledgement is broader than citation.

You cite when you quote, paraphrase or reproduce specific AI-generated content. You acknowledge when AI shaped your work in ways that do not produce a quotable passage — brainstorming an essay structure, suggesting search terms, rewriting your own sentences.

Most institutions now want a short statement, usually in the methods section or a note at the end, saying what you used and what for. Something like:

ChatGPT (GPT-4o, OpenAI) was used to suggest an initial structure for this essay and to identify counter-arguments. All research, analysis and writing are my own.

That sentence takes thirty seconds and is the difference between transparency and a misconduct hearing. It also protects you: a marker who suspects AI involvement and finds it declared will read the work quite differently from one who finds it concealed.

Three practical warnings

Do not cite an AI for a fact. If a model tells you something true, the citation belongs to whatever source actually establishes it. Cite the AI only for text you are reproducing as AI output — otherwise you are citing the messenger.

Check every reference the model gives you. Fabricated citations remain common: real authors, real journals, plausible titles, and a DOI that resolves to something else entirely or to nothing. Verify each one against the actual source before it enters your bibliography. A DOI you can resolve is real; a DOI you have not checked is a guess.

Versions move. Model names change and old versions are withdrawn. Record the model and the date at the moment you use it, because you will not be able to reconstruct them later.

Keeping the reference tidy

Once you have the details, the mechanical part — the italics, the stop placement, the order of the elements — is what a generator is for. Our citation generator handles APA 7, MLA 9, Chicago and Harvard, and imports from RIS or BibTeX if you already have references in a manager.

One caution that applies to every citation generator including ours: automatic sentence case cannot reliably tell a proper noun from an ordinary one. Read your finished references before submitting, and check the capitals.

Sources

The two primary authorities, both freely readable, and both worth checking directly since this guidance is still being revised:

• APA Style: How to cite ChatGPT
• MLA Style Center: How do I cite generative AI in MLA style?