How to Cite ChatGPT and Other AI Tools: APA, MLA, Chicago, and Harvard Examples

Decide which case you are in before you format anything. A citation shows where words or ideas came from. A disclosure explains how you used a tool. A fact from an AI answer is only a lead to the real source.

For how to word a disclosure, see our guide to AI and academic integrity. In every case, your instructor's or journal's rules come first.

All four examples cite one invented chat. On September 28, 2026, a student gives ChatGPT (model GPT-5) the prompt "List the main arguments for and against later start times for high schools." The student renames the chat "Arguments about later school start times" and shares it.

APA updated its guidance in September 2025. The APA Style blog now lets you cite the specific chat when that helps readers. Most AI chat tools, it notes, can share a chat with its own URL and title. The 2023 post with the "ChatGPT (Mar 14 version)" format is kept for archival purposes only. A chat reference has four parts:

• Author: the company behind the tool, such as OpenAI or Google. APA says AI itself cannot be an author.
• Date: the year, month, and day of the chat.
• Title: the chat title in italic sentence case, then [Generative AI chat]. Rename the chat first if its title says little.
• Source: the tool or model name, then the chat URL.

Keep the prompt out of the reference. Describe it in your sentence and cite the response: Asked for arguments about later high school start times, ChatGPT said later starts "could better match teenagers' sleep cycles" (OpenAI, 2026). If the full prompt matters, APA suggests the Method section or an appendix, as agreed with your instructor or editor. Paraphrased output follows the usual paraphrase rules.

If AI only edited or translated your own writing, APA cites the tool in general, not each chat. The year is the one in which the tool was last updated. APA's own example:

OpenAI. (2025). ChatGPT [Large language model]. https://chatgpt.com/

MLA does not treat the AI tool as an author. Its revised guidance (2025) fills the core elements like this:

• Title of Source: a description of what was generated, such as the prompt in quotation marks followed by the word prompt.
• Title of Container: the tool, in italics.
• Version: the model, as specifically as possible.
• Publisher: the company that made the tool.
• Date: the day the content was generated.
• Location: the share link, or the tool's general URL if there is none.

In your text, use the first words of the title: ("List the main arguments"). MLA also allows a more general title, such as Later school start times prompt, since a conversation can run through several prompts.

Chicago

The Chicago Manual of Style Q&A says to credit ChatGPT whenever you use text it generated. For most writing, naming the tool in your sentence is enough. A student paper or research article may need a numbered note with the tool, company, date, and URL. Add the prompt to the note if your text does not give it. If you edited the output, say so.

Georgetown University Library quotes section 14.112 of the 18th edition: chatbot conversations are not usually included in a bibliography.

Harvard

Harvard varies by university, so follow your own library's guide. UCL's guide is one example. UCL normally asks for an Acknowledgements section and no reference list entry. Where a reference is needed, for example because your work discusses AI output or your department requires it, UCL treats the output as a work with no author, as in the table above.

Gemini, Claude, Copilot, and other text generators follow the same pattern. The company is the author in APA and the publisher in MLA. Record the model as well as the tool. MLA puts the model in the Version element. APA no longer asks for a version number by default. It suggests the model name, such as Gemini 2.5 Flash, if one is available. The tool name alone is also acceptable.

APA tells researchers to document their prompts for their own records and compares this to keeping a full data set. A saved record also helps if a link stops working.

• Save the full conversation as text or PDF, with the date and the model.
• Keep the chat. Deleting a ChatGPT conversation removes its share link, and deleting your account removes your share links (OpenAI Help Center).
• A share link from a personal account is a snapshot. Later messages appear only after you update the link.
• Share links from ChatGPT Business, Enterprise, or Edu workspaces open only for eligible members of that workspace.

AI answers can contain wrong facts and invented references. Walters and Wilder (2023) had ChatGPT-3.5 and ChatGPT-4 write short literature reviews on 42 multidisciplinary topics, 84 papers in all, and checked the 636 references in them. Of the GPT-3.5 references, 55% were fabricated, compared with 18% of the GPT-4 references. Among the real references, 43% from GPT-3.5 and 24% from GPT-4 had substantive citation errors.

MLA recommends consulting and citing the sources an AI links to instead of the AI response, and warns that AI tools can make up sources. APA likewise says to read those sources yourself and cite them.

Worked example

Say a chatbot claims that a 2019 study found 12 percent higher attendance after a later school start. The claim is invented for this example.

1. Search your library catalog or a scholarly database for the study.
2. If it exists, read it and check the number and its context.
3. Cite the study, for example (Author, 2019), not the chatbot.
4. If you cannot find it, drop the claim or find a published source.

If you only reach a study through another publication, see how to cite a secondary source.

A similarity check compares your draft with the sources it can find. Use it to confirm that borrowed passages still carry citations after editing. It does not show whether text came from an AI, and it may miss a chat you quote. According to OpenAI, shared chat pages are not intended for search-engine indexing. Check quoted AI output against your saved transcript.

The Certentis plagiarism checker marks the overlaps it finds, links each one to its source, and shows a similarity score. Similarity is not automatically plagiarism, so you judge each passage, as our guide to understanding similarity reports explains. A report may not find every source, and Certentis publishes no accuracy figure for the checker yet (benchmarks).