We think about ChatGPT citations from two angles at OTH3L: how to format them correctly in academic work, and how brands earn citations from AI search tools. This one is about the first.
Why ChatGPT Doesn’t Follow Normal Citation Rules
Two people can ask ChatGPT the same question, on the same day, at the same time and still both will get different answers. A conversation from March won’t regenerate the same output in June, even with the exact same prompt. There’s no traditional author, page number, or archived version anyone else can pull up and check.
That’s why APA solves this by assigning authorship to OpenAI. MLA refuses to call LLMs an author and uses the prompt description as the title instead. Meanwhile, Chicago sidesteps the author question completely and keeps the entire thing in a footnote.
APA vs. MLA vs. Chicago: The Differences That Actually Matter
| APA | MLA | Chicago | |
|---|---|---|---|
| Treats AI as | A corporate author (OpenAI) | Not an author at all; the prompt becomes the title | Not an author; handled as a footnote source |
| Last major update | September 9, 2025 | August 13, 2025 | CMOS 18th edition Q&A; no version-stamped updates |
| Format depends on | Whether you’re citing a specific chat or the tool generally | The exact prompt you use | Whether the prompt already appears in your text |
| Needs a public URL? | Preferred for specific chats; not required for general tool citations | Preferred — a shareable conversation link | Only if you want a bibliography entry; otherwise footnote only |
| In-text citation | (OpenAI, 2026) | Shortened prompt in quotes | Numbered footnote; no parenthetical |
If an article you’re reading doesn’t mention a 2025 update for both APA and MLA, it’s working from outdated guidance.
How to Cite ChatGPT in APA (Current as of September 2025)
APA replaced the single template from 2023 with two separate ones, and which one you use depends on what you actually did with the chat.
Use Template A when the chat generated ideas, content, or information you’re drawing from in your paper.
OpenAI. (2026, March 3). Vanity metrics in startup board decks [Generative AI chat]. ChatGPT. https://chatgpt.com/share/[your-chat-id]
The title in brackets is new. The 2023 format never asked you to describe what the chat was about — it just cited “ChatGPT” as a generic tool. The 2025 update treats each chat as its own source with its own title, which makes a reference list with multiple ChatGPT citations actually distinguishable from each other.
Use Template B when you used the tool generally, like editing your own writing, and you’re not pointing to one specific conversation.
OpenAI. (2026). ChatGPT [Large language model]. https://chatgpt.com/
In-text, both templates collapse to the same short form: parenthetical (OpenAI, 2026), or narrative OpenAI (2026) — the same way you’d cite a report from a government agency.
One thing you should stop: hunting down a specific version number like “(Mar 14 version).” That’s archived now. APA’s current preference is the model name if you have it, and chasing a version is not required.
If you used ChatGPT for research help, APA also wants this documented — usually in your methods section: what you asked it, what it gave you, and whether you checked the output before using it. That’s separate from the citation itself.
How to Cite ChatGPT in MLA (Current as of August 2025)
MLA’s core decision — made back in 2023 — is unchanged: AI tools don’t get treated as authors. Instead, a description of what the AI produced goes in the Title of Source slot.
"Summarize the main arguments against using vanity metrics in startup board decks" prompt. ChatGPT, GPT-5, OpenAI, 3 Mar. 2026, chatgpt.com/share/[your-chat-id].
The prompt goes in quotation marks, the tool name is italicized, then the model name, the publisher, the date, and the link.
The August 2025 update changed two things: First, a stable, shareable link to the actual conversation is now the preferred URL, not the general domain. Second, MLA now wants the specific model name (e.g. GPT-5.5) in the version slot.
MLA also draws a line that APA doesn’t bother with: citing AI output and disclosing AI use are two separate jobs. Citing covers the mechanics of what you used. Disclosing means explaining what role it played in your actual work. Some instructors want a full transcript of your prompts attached as an appendix.
One more important rule: if ChatGPT points you toward a source, click through and cite that original source directly. Don’t cite ChatGPT’s summary of someone else’s work as if it were the work itself.
How to Cite ChatGPT in Chicago
Chicago is the simplest of the three, and it hasn’t needed a major overhaul because the footnote-only approach already handled the “this isn’t a stable document” problem from the start.
1. Text generated by ChatGPT, OpenAI, March 3, 2026, https://chatgpt.com/.
If you’ve already stated the prompt in your body text, the footnote can skip it. If you haven’t, add it: ChatGPT, response to “summarize the arguments against vanity metrics,” March 3, 2026, OpenAI.
Rule that trips people up: don’t put ChatGPT in your bibliography unless you can provide a public URL that doesn’t require someone else’s login to access. Most ChatGPT conversations don’t qualify, which is why Chicago treats this closer to a personal communication — like an email or a phone call — than a citable document.
Why Half of ChatGPT’s Citations Don’t Exist
Across six peer-reviewed studies covering 732 total citations, 51% were fabricated. Individual study rates ran from 17% up to 94% depending on the discipline and the model version. As models advance, the error rate is dropping — but it’s still nowhere close to safe to use unverified.
The Verification Workflow That Actually Catches Fake Sources
This takes less than five minutes per source and doesn’t require any institutional database access.
- Search the full title in Google Scholar, in quotation marks.Paste the exact title into scholar.google.com. You’ll see a real article with publisher links or library access options. No results is a red flag.
- Run the DOI through Crossref or doi.org.Crossref’s metadata search is public with no login required. If the DOI doesn’t resolve, or resolves to a completely different article, the source is highly suspicious — avoid it.
- Confirm the journal itself is real.Search the journal name on its own. If there’s no official website, ISSN, and indexed presence anywhere, skip it.
A source missing from Google Scholar doesn’t automatically mean it’s fake — smaller or newer journals have uneven indexing. But if a source is missing from all independent databases, that’s a clear signal to avoid it.
Is Using AI Allowed?
There’s no universal yes or no. Policy varies with whoever is setting the rules for that specific piece of writing: an instructor, a department, a publisher, an editor. Harvard, MIT, and Stanford all set policy at the school or instructor level.
This habit will save you: note down where you used ChatGPT, verify every source it cites before it goes in your reference list, and check your actual syllabus before assuming anything is or isn’t allowed. “AI tools were used” as a disclosure isn’t specific enough for most institutions. You need to name the tool, the model, the date, and exactly what you used it for.
Conclusion
To cite ChatGPT: APA now uses two templates (a specific chat, or the tool generally) as of its September 2025 update; MLA puts your prompt description in the title slot as of its August 2025 update; and Chicago uses a footnote with no bibliography entry unless there’s a public URL. All three capture the date and model.