If you used an AI tool in your work and your institution permits it, you have to say so. All three major style guides have now published guidance, and they arrived at genuinely different answers - because they disagree about what a conversation with a language model actually is.

That disagreement is worth understanding, because it tells you what to do in the cases the templates do not cover.

The underlying problem

A citation exists so a reader can find the source. That assumption breaks completely with a chatbot.

Ask the same model the same question twice and you may get different answers. Ask it next year and the model has been replaced. Your reader cannot retrieve what you saw, cannot verify you represented it fairly, and cannot check whether it was accurate. Nothing in the citation machinery was designed for a source that only existed once, for you.

Each style guide resolved this differently, and each resolution is defensible.

APA: treat it as software

APA treats the model as an algorithm you used - closer to SPSS than to an author. The company is the author, the model is the work, and the version is what makes it identifiable.

Reference list:

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

In text: (OpenAI, 2026)

APA's guidance is explicit that because the output is not retrievable, the prompt and the response should be included in your text or an appendix if they matter to your argument. That is the important part and the part people skip. The reference tells your reader which tool; the appendix is what lets them evaluate what it actually produced.

Substitute the relevant company and model for other tools - Anthropic for Claude, Google for Gemini - and keep the version, because it is the only thing distinguishing what you used from what exists now.

MLA: treat it as a source you consulted

MLA takes a different view. Its template has no fixed author slot for AI, and MLA explicitly recommends not treating the model as an author. Instead the prompt goes in the title position, because the prompt is the thing that identifies this particular output.

Works Cited:

"Explain the causes of the 1929 crash in three paragraphs" prompt. ChatGPT, Aug 2026 version, OpenAI, 3 Aug. 2026, chat.openai.com.

In text: ("Explain the causes")

This is the more informative format, because the prompt is right there in the reference. It is also more cumbersome, especially if you used a long or iterative prompt - in which case MLA suggests a shortened form in the entry and the full prompt in a note or appendix.

Chicago: a note, usually not a bibliography entry

Chicago's position is the most pragmatic. Because the content is not retrievable, it generally does not belong in a bibliography - a bibliography is a list of things a reader could go and read. Chicago puts it in a footnote instead, and is comfortable with the citation being conversational.

Footnote:

Text generated by ChatGPT, OpenAI, August 3, 2026, chat.openai.com.

Chicago also accepts crediting it in the text alone - "when prompted about the causes of the 1929 crash, ChatGPT suggested..." - with no formal citation at all, provided the prompt is described. For a single passing mention that is often the cleanest option.

Which one applies to you

Your department's rules override all of this, and many now have their own AI-use policy that is more specific than any style guide. Check that first - the style guides govern formatting, but your institution governs whether and how you may use these tools at all.

Where you have a choice, the styles behave as they always do: APA identifies the tool, MLA identifies the specific output, Chicago keeps it out of the way.

The part that matters more than formatting

Formatting a citation is the easy half. Two other things carry real weight.

Disclose the actual use, not a vague one. "AI was used in preparing this document" tells a reader nothing. There is a large difference between using a model to brainstorm angles, to check grammar, to draft a paragraph you then rewrote, and to produce text you submitted largely unchanged. Many institutions now require a short statement describing which of these applied and where. Write it specifically, because a vague disclosure reads as an attempt to cover everything without admitting anything.

Verify every factual claim independently. Language models produce confident, well-formed, entirely fabricated citations - papers that do not exist, with plausible authors, in real journals, with DOIs that resolve to something else. This is not an occasional glitch; it is a predictable consequence of how the systems work. If a model gave you a reference, find the actual paper before citing it. If you cannot find it, it may well not exist.

That failure mode has already produced sanctions in courts and retractions in journals, and the pattern is always the same: someone cited a source they never opened.

Cite what you actually read. If a model summarised a paper and you cite the paper, you are claiming to have read it. If you then rely on a detail the summary got wrong, that is your error and it will look like one. Read the source, or cite the model.

A note on where this is heading

All three style guides have revised their AI guidance more than once, and will again. Anything you read on this - including this post - should be checked against the current official guidance before you submit something that matters. APA's style blog, the MLA Style Center, and the Chicago Manual's Q&A all publish updates directly, and are the authoritative versions.

For the sources that are stable and retrievable - journal articles, books, web pages - our Citation Generator will format them in APA, MLA, Chicago, Harvard, IEEE or Vancouver from a DOI, ISBN, URL or title. It deliberately does not generate chatbot citations, because the one thing that makes those citations useful is the prompt and context you supply, and no tool can produce that for you.

How to Cite AI-Generated Content in APA, MLA and Chicago