Papex
Explore
← Blog

Guide21 min read

How to cite ChatGPT, Claude AI and Gemini in APA, MLA and Chicago

APA 7, MLA 9, Chicago, AMA, IEEE and Harvard formats for ChatGPT, Claude and Gemini, from each guide's current advice, and when to disclose instead of cite.

Papex

To cite ChatGPT, Claude or Gemini, APA names the company that makes the tool as the author, MLA starts with a description of your prompt, and Chicago uses a numbered note. In all three, link to a shared copy of the chat where you can. First check your journal's policy, though: many treat AI output as something to disclose in the methods or acknowledgements, not as a source for the reference list.

Every format below follows the style guide's own published guidance, checked 9 October 2026, except Harvard, where we use university guides based on Cite Them Right. Where a guide shows only ChatGPT, we have applied its pattern to Claude and Gemini and say so.

Should you cite an AI chat or disclose it?

Citing and disclosing do different jobs. A citation points readers to the source of words or ideas you used. A disclosure tells them how a tool took part in the work: which tool, what it did and what you checked. A paper may need one, both or neither.

The MLA puts the difference plainly in its post Beyond Citation: Describing AI Use in Your Work:

"While a generative AI tool may direct you to a source and, in the case of some AI tools, even include snippets or excerpts from a source, the AI output itself is not a source."

APA's style team describes a different arrangement for APA's own journals: disclose the use, and also cite the tool in the reference list (APA Style, Part 3). So whether you cite depends on what the AI did and on who will publish or grade the work.

What the AI didCite it?Disclose it?
It produced text you quote or paraphraseYes, unless your journal bars AI as a referenceYes
It helped you find papersNo: cite the papers you readOnly if it was part of a formal search strategy (APA)
It checked spelling or grammar inside your word processorNoNo
A chatbot edited, restructured or translated your textAPA: cite the tool in generalUsually; journals differ on language-only edits
It wrote or fixed analysis codeAPA: cite the tool, or the chat if the code is centralYes, in the methods
It gave you a fact or a referenceNo: find and cite the originalIf it shaped the work

The rows come from APA Part 1 and Part 2, the MLA's revised guidance, and the Elsevier and Sage policies below. For what a disclosure statement should say and where it goes, see How to disclose AI use in a research paper.

What do journals say about citing AI as a source?

Policies differ, and your journal's instructions come before any style guide. Here is a sample, checked 9 October 2026:

PolicyOn citing AIOn disclosing AI
ICMJE Recommendations (followed by many medical journals)"Referencing AI-generated material as the primary source is not acceptable."In the cover letter and in the submitted work
SageCite original sources "rather than GenAI tools as primary sources"; citing a tool may suit AI-made imagesIn the methods or acknowledgements; language, grammar and structure edits are exempt
ElsevierDo not list AI tools as authors, "nor cite AI tools as an author"A separate AI declaration just before the references; research use goes in the methods
Taylor & FrancisNot covered by the policyA statement giving the tool's full name and version number, how it was used and why
JAMA NetworkNot stated; AI must not be used to generate or format referencesIn the Acknowledgment, or the Methods if part of the study design
APA Journals, as summarised by APA StyleCite the tool, with a reference-list entryIn the Method section, author note or another relevant place

APA and MLA don't treat the tool as an author, and ICMJE, Elsevier and Taylor & Francis rule out listing it as one. If your journal follows the ICMJE or Sage line, leave the chat out of the references. Instead, quote what you used in the text, describe the use where the journal asks, and put the transcript in the supplementary material.

How to cite ChatGPT, Claude and Gemini in APA 7

APA replaced its ChatGPT advice in September 2025 with a three-part series and a page of AI reference examples. There are now two reference types: one for a specific chat, and one for the tool in general. In both, the author is the company: OpenAI for ChatGPT, Anthropic for Claude and Google for Gemini.

In the examples below, xxxx stands for the ID in your own share link and [name] for the model your app shows. Everything else is literal, including APA's square-bracketed descriptions.

Citing a specific chat in APA

APA's template is:

AI Company Name. (year, month day). Title of chat in italics [Description, such as Generative AI chat]. Tool Name/Model. URL of the chat

The date is the day the chat took place or ended. The title is the chat's own title, in italics and sentence case. Most apps let you rename a chat, and APA suggests giving it a descriptive title before you cite it. Use this form when seeing the chat will help readers.

In the text, write (OpenAI, 2026), (Anthropic, 2026) or (Google, 2026), or narratively, OpenAI (2026). The source slot can hold the model instead of the tool name, as in APA's own example in Part 1 of its series:

Anthropic. (2025, May 20). Essential grammar topics for high school graduates [Generative AI chat]. Claude Sonnet 4. https://claude.ai/share/329173b2-ec93-4663-ac68-4f65ea4f166d

The prompt does not go in the reference. APA says to give it in the text or, if it is long, in an appendix or supplementary material. A methods sentence might read: "We used ChatGPT (OpenAI, 2026) to draft R code for the mixed-effects models and checked each line before running it; the prompts are in Appendix B."

Citing the tool in general in APA

APA's template is:

AI Company Name. (year). Tool Name/Model in Italics and Title Case [Description; e.g., Large language model]. URL of the tool

Cite the tool in general when citing a specific chat would be "unhelpful, unethical, or otherwise inappropriate". APA's examples include using AI to edit or translate your own writing, and chats with study participants, where one reference per chat could compromise confidentiality. The year is the year the tool was last updated. If the tool doesn't say, use the copyright year on its site or app.

Is the "(Mar 14 version)" format still correct?

No. If a guide shows "OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model]. https://chat.openai.com/chat", it is copying APA's 2023 post, which APA now labels "retained for archival purposes only". APA no longer advises a version number by default. Name the model if you can, or just the tool. If a tool does show a version, put it in parentheses, not in italics, after the tool or model name.

Chatbot output is not a personal communication in APA either. Even the 2023 post said that "with ChatGPT-generated text there is no person communicating."

How to cite ChatGPT, Claude and Gemini in MLA 9

The MLA's revised guidance (August 2025) does not treat the tool as an author. The entry starts with a description of what was generated, usually your prompt. The tool is the container, and the Version slot holds the model, named "as specifically as possible". Then come the company, the date the content was generated, and a stable share link. If the tool has no share link, give its general URL instead.

The MLA's own example, exactly as published:

“Describe the theme of nature in Jane Austen’s Mansfield Park” prompt. ChatGPT, model GPT-4o, OpenAI, 23 Sept. 2024, chatgpt.com/share/66f1b0a0-d704-8000-be9a-85f53c850607.

The MLA's examples cover ChatGPT, DALL-E and Copilot only. Applied to all three chatbots, the pattern gives:

  • “Explain the difference between fixed and random effects” prompt. ChatGPT, model [name], OpenAI, 6 Oct. 2026, chatgpt.com/share/xxxx.
  • “Explain the difference between fixed and random effects” prompt. Claude, model [name], Anthropic, 6 Oct. 2026, claude.ai/share/xxxx.
  • “Explain the difference between fixed and random effects” prompt. Gemini, model [name], Google, 6 Oct. 2026, g.co/gemini/share/xxxx.

The in-text citation uses the first words of the description: (“Explain the difference”). MLA's examples drop the "https://" from links. If the chat went through several prompts, the MLA accepts a shorter, general description, such as: Fixed and random effects prompt.

Citation is only part of it for the MLA. It also advises acknowledging substantive AI use beyond citation, and its 2023 guidance asked writers to acknowledge editing or translating "in a note, your text, or another suitable location".

How to cite ChatGPT, Claude and Gemini in Chicago style

Chicago's Q&A answer on AI relies on the 18th edition. It says to credit the tool whenever you use text it generated. For most writing, naming the tool in your sentence is enough. For a student paper or research article, use a numbered note, with the URL copied from the app's Share button. Chicago's example is "Text generated by ChatGPT, OpenAI, January 10, 2026, URL." Following the same pattern:

  1. Text generated by ChatGPT, OpenAI, October 6, 2026, https://chatgpt.com/share/xxxx.
  2. Text generated by Claude, Anthropic, October 6, 2026, https://claude.ai/share/xxxx.
  3. Text generated by Gemini, Google, October 6, 2026, https://g.co/gemini/share/xxxx.

If your text doesn't give the prompt, put it in the note. The comma goes inside the closing quotation mark:

  1. Claude, response to “Explain the difference between fixed and random effects,” Anthropic, October 6, 2026, https://claude.ai/share/xxxx.

If you edited the output, say so in the text or at the end of the note, for example "edited for style and content". Small changes such as switching to curly quotes don't need a mention.

What about a bibliography entry or author-date style? The current answer shows notes only. An earlier version, based on the 17th edition (archived June 2025), gave the author-date form "(ChatGPT, March 7, 2023)". It also said to leave AI chats out of the bibliography unless you provide a publicly available link. The revised answer drops both points, and the Manual itself is subscription-only. If your publisher wants a bibliography entry or an author-date citation, ask the editor or check the 18th edition.

How to cite ChatGPT, Claude and Gemini in AMA style

AMA style, which JAMA and the JAMA Network journals use, gives no reference-list format for an AI chat in its public guidance. You acknowledge the use instead: in the Acknowledgment section, or in the Methods if the AI was part of the study design.

The AMA's blog, AMA Style Insider, gives this example from section 3.15.13 of the AMA Manual of Style:

"The authors acknowledge using ChatGPT (GPT-3.5, OpenAI) for text editing to improve the fluency of the English language in the preparation of this manuscript on September 15, 2023." It goes on to say that the meaning is unchanged, that ChatGPT did not shape the intellectual content, and that the authors take full responsibility.

JAMA's instructions for authors ask for the tool's name, version and extension numbers, manufacturer, dates of use, how it was used and on which parts, and confirmation that the authors take responsibility for the content. Basic grammar and spelling checkers are exempt. If AI was part of the study, the Methods should also give the prompts, their order and any revisions. Applying AMA's pattern to all three:

  • The authors acknowledge using ChatGPT ([name], OpenAI) on October 6, 2026, to suggest edits for clarity to the Discussion. The authors reviewed every change and take full responsibility for the integrity of the content.
  • The authors acknowledge using Claude ([name], Anthropic) on October 6, 2026, to …
  • The authors acknowledge using Gemini ([name], Google) on October 6, 2026, to …

JAMA adds that AI "should not be used to generate or format references". Section 3.15.13 itself is subscription-only, so we could not read it; Butler University Libraries quotes it as saying that descriptions of AI use "should appear in running text and not be included in the formal reference list".

How to cite AI in IEEE style

IEEE has no reference format for AI chats. Its Reference Guide (version 3.28.2025) covers blogs, datasets, software, websites and more, but has no entry for ChatGPT, Claude, Gemini or any other AI tool. We have not invented one.

IEEE's ethics principles (page 5) say an AI tool "may be cited in references and bibliographies but only used as a source if the author has independently verified the accuracy of the content". So a reference is allowed, but IEEE has not said what it should look like. If your editor wants one, ask which form they prefer.

What IEEE does set out is an acknowledgement. Its author policy says AI-generated content, "including but not limited to text, figures, images, and code", must be disclosed in the acknowledgments section. Name the AI system, say which sections use AI-generated content, and explain briefly how much the AI did. Using AI only to edit language and grammar is "generally outside the intent" of the policy: disclosure is "not required, but recommended". IEEE also advises leaving the reference list out when you run text through an AI editor, so the tool can't make unwanted changes.

In IEEE style the heading is singular, "Acknowledgment", with no "e" after the "g" (IEEE Editorial Style Manual for Authors). IEEE gives no wording for text, so these follow its three requirements:

  • The authors used ChatGPT ([name], OpenAI) to draft the Python code in Section III-B. The authors tested and revised the code.
  • The authors used Claude ([name], Anthropic) to draft the Python code in Section III-B. The authors tested and revised the code.
  • The authors used Gemini ([name], Google) to draft the Python code in Section III-B. The authors tested and revised the code.

How to reference AI in Harvard style

Harvard has no single official version. Universities publish their own, so your university's guide comes first. Universities such as Greenwich, Newcastle and De Montfort base theirs on Cite Them Right (Bloomsbury). Its AI page links to an entry on referencing generative AI, but that entry needs an institutional subscription, so we could not read it. The formats below come from university guides that say they follow it, checked 9 October 2026.

The University of Greenwich's guide (September 2026) takes its Harvard examples from Cite Them Right Online. It names the AI company as author:

Name of AI company (Year) AI-generated [medium] by [name of AI] with prompt '…', Day month. Available at: URL of chat (Accessed: day month year).

Applied to the three chatbots:

  • OpenAI (2026) AI-generated text by ChatGPT with prompt 'Explain the difference between fixed and random effects', 6 October. Available at: https://chatgpt.com/share/xxxx (Accessed: 6 October 2026).
  • Anthropic (2026) AI-generated text by Claude with prompt 'Explain the difference between fixed and random effects', 6 October. Available at: https://claude.ai/share/xxxx (Accessed: 6 October 2026).
  • Google (2026) AI-generated text by Gemini with prompt 'Explain the difference between fixed and random effects', 6 October. Available at: https://g.co/gemini/share/xxxx (Accessed: 6 October 2026).

In the text: "Claude (Anthropic, 2026) suggested…" or "(Anthropic, 2026)".

If only you can see the chat, Newcastle University's style sheet (2025) treats it as a personal communication and describes the material in the in-text citation. Its example is "OpenAI ChatGPT (2023) ChatGPT response to Lorna Smith, 3 April." On that pattern: Anthropic Claude (2026) Claude response to [your name], 6 October.

The two guides differ, so check which version your course uses, and whether it wants the full chat as an appendix.

How to cite an AI-generated image

First check that the image is allowed. JAMA does not accept clinical images or illustrations created or altered by AI unless they are part of the formal research design or methods (instructions). Elsevier allows AI help with explanatory images such as flow charts, disclosed in each caption, but not for creating or altering primary research images such as microscopy or patient images.

MLA puts the citation in the caption, or the same details in the works-cited list. Its example:

Fig. 1. “Create an expressionist-style image of two people standing on a beach looking at the ocean” prompt, DALL-E, version 3, OpenAI, 23 Sept. 2024, chatgpt.com/share/66f1c3a3-3f90-8000-9750-82c57c4a6592.

For images made in ChatGPT or Gemini:

  • Fig. 1. “Draw a plant cell with the organelles labelled” prompt, ChatGPT, model [name], OpenAI, 6 Oct. 2026, chatgpt.com/share/xxxx.
  • Fig. 1. “Draw a plant cell with the organelles labelled” prompt, Gemini, model [name], Google, 6 Oct. 2026, g.co/gemini/share/xxxx.

Claude doesn't generate photos or illustrations, but it can build diagrams and charts (Claude Help Center); cite those the same way, with Claude, Anthropic and claude.ai/share/xxxx.

APA asks you to "indicate that you used AI in the table or figure note" and to cite the tool in general (Part 1). It gives no separate image reference. A note might read: Note. Created with Gemini (Google, 2026); the prompt is in Appendix A. The reference is the general tool reference shown earlier. AI features inside common software, such as Canva, need no citation (Part 2).

IEEE's style manual gives this caption: "Graphic(s) created using AI-generation. For image credits, please see the Acknowledgment section of this article." The Acknowledgment then reads "Fig. X was created using <AI system used>.", followed by a brief explanation of how far the AI was used.

How do I give the exact model and version?

Note what the app's model picker shows on the day you use it, and take a screenshot. APA suggests asking the AI which model it is. If you do, compare the answer with the picker, because AI output can be wrong.

  • APA: put the tool or model name in the title (tool reference) or the source slot (chat reference). Add a version, in parentheses and not in italics, only if the tool shows one.
  • MLA: put the model in the Version slot, for example "model GPT-4o".
  • AMA: name the tool, then the model and manufacturer in brackets, as in AMA's own example, ChatGPT (GPT-3.5, OpenAI). JAMA also asks for version and extension numbers.
  • IEEE: name the AI system in the Acknowledgment.
  • Chicago: the Q&A's notes name only the tool. If the model matters, name it in your sentence or your disclosure statement.
  • Disclosure statements: Elsevier asks for the model or tool, the version and the developer when AI is part of the methods. Taylor & Francis asks for the tool's full name with its version number.

What to keep so others can check your citation

The same prompt can give a different answer in each chat. APA made the point in 2023: "ChatGPT will generate a unique response in each chat session, even if given the same prompt." A reference only helps if the exact chat can still be found. Keep these:

  • The share link, and its limits. In ChatGPT personal accounts, Claude and Gemini, a shared link is a snapshot: messages added later are not included unless you update or re-share it (OpenAI, Anthropic, Google). A Claude snapshot leaves out attached files. Links made in ChatGPT Business, Enterprise or Edu workspaces, and Claude Team or Enterprise chats, open only for members of that organisation, so a journal's readers cannot follow them.
  • A copy that outlives the link. Deleting a ChatGPT conversation removes its shared link. A Gemini public link is deleted if any part of the chat is removed from your Gemini Apps Activity, including by auto-delete. Claude offers an export under Settings > Privacy, and the download link expires 24 hours after it arrives. ChatGPT personal accounts also offer a data export. For Gemini, copy the chat into your own files.
  • Every prompt, in order, including the follow-ups. APA says researchers "should document the prompts used with AI tools for their own records".
  • The model name and the date, with the picker screenshot.
  • The exact output you used, and what you changed.

When AI is part of a study, for example a chatbot talking to participants, APA advises keeping those chats "just as you would a full data set".

Keeping the record as you go

A share link records a chat. It doesn't record which version of your paper relied on that chat, or when that version existed. That second part is what Papex keeps.

In Papex, a hypothesis, finding or publication is a record with versions. When you sign a version, the signature links you to that exact version. A later edit becomes a new version with its own signature and date, and earlier versions stay as they were. Only a salted fingerprint of the version leaves Papex. It is timestamped with OpenTimestamps and anchored in Bitcoin, meaning a commitment to it is recorded in a Bitcoin block that anyone can check. The .ots proof works with any OpenTimestamps tool, without Papex (what a Bitcoin timestamp proves). After signing, you can keep a version private, share it with people you invite, or publish it.

With your permission, Claude, ChatGPT, and tools such as Claude Code, Codex and Cursor (through the Papex MCP package) can prepare private drafts, propose changes and look up references. They cannot sign, accept their own proposals, make rights statements or publish; only you do that, and an AI app is never listed as an author. Model and tool names are kept as the AI agent reports them and labelled "Agent-reported"; Papex does not verify them. Papex receives only what an AI app sends in a tool call, not your conversation, so keep your share link and export too.

A signed record shows who signed which exact version, and that the version existed by a date. It does not show that your citation is correct, that the AI output was right, or that a journal will accept the work.

To see how AI apps connect, read /docs/ai.

Sources

Keep reading

How to write an AI disclosure statement for a research paper

AI disclosure statement examples and templates for research papers: what to include, where it goes, if grammar help counts, and 11 journal policies compared.

Guide

How to protect a research idea from being scooped

Afraid of being scooped? You can't own an idea, but you can show what you had and when. Preprints, preregistration, notebooks and timestamps compared.

Guide

How OpenTimestamps works: what a Bitcoin timestamp proves

What an OpenTimestamps proof shows about your work and what it does not, how exact the Bitcoin date is, and how to verify an .ots file yourself.

Engineering