How to Cite AI Conversations in Academic Work: APA, MLA, Chicago, and Why History Matters

If you used ChatGPT, Claude, Gemini, or another AI tool while writing a paper, completing a research project, or drafting a thesis, the question of how to cite that use is no longer theoretical. Universities, journals, and individual instructors all have policies now. Most of those policies have converged around two simple expectations: disclose AI use, and be able to evidence it if asked.

This guide covers what those expectations look like in practice — how current citation styles treat AI tools, what the underlying integrity issues are, and how to preserve AI conversations so that you can actually meet the evidence expectation months after the conversation happened.

Why AI citation became a thing so quickly

Academic citation exists for two reasons: to credit the source of an idea, and to make the work reproducible by giving readers a path back to that source. AI conversations sit awkwardly against both.

Credit. A conversation with ChatGPT is not the work of a single identifiable author. It is the output of a model trained on a corpus the model's creators may not fully disclose. Citing "ChatGPT" credits the tool, not the underlying ideas, and not the humans whose work the model was trained on.

Reproducibility. A second person running the same prompt at a later date may get a different output. Model versions change. Internal sampling is non-deterministic on many platforms. The thing being cited is therefore harder to verify than a published source.

Both issues mean AI citation is partial by nature — it documents that AI was used and what it returned, not who originally said the thing and not what someone else would have got from the same prompt. The point of the citation is transparency and provenance for the work the author produced, not equivalence to a normal source citation.

This is why the disclosure expectation is sometimes stronger than the citation format itself. The institution wants to know AI was used; the format is a way of recording it.

APA: current shape of the guidance

APA's published guidance has converged on treating AI outputs as software citations with author-as-organisation. The typical shape:

OpenAI. (2026). ChatGPT (Mar 14 version) [Large language model]. https://chat.openai.com

With an in-text citation that names the prompt used:

When prompted to "summarise the key findings of [paper X]," ChatGPT (OpenAI, 2026) returned…

Important specifics in current APA guidance:

• AI is cited as the author (organisation), not as a personal communication.
• The model version and date matter — different versions produce different outputs.
• The prompt should be reported in-text or in an appendix.
• The AI output should be preserved (the URL alone is not sufficient because outputs are non-deterministic).

APA updates this guidance periodically. Check the APA Style Blog for the current canonical format before submitting.

MLA: focus on the prompt and the response

MLA's approach centres on documenting the prompt-response interaction explicitly. A typical MLA-format Works Cited entry:

"Summarize the key findings of [paper X]" prompt. ChatGPT, 14 Mar. 2026 version, OpenAI, 14 Mar. 2026, chat.openai.com/chat.

In-text references typically use a short form referring to the AI tool and date.

Key MLA-specific notes:

• The prompt is treated as part of the citation.
• Version and date are required.
• MLA explicitly recommends that the writer disclose how AI was used in a methods or process note, not just in the citation.

See the MLA Style Center guidance on generative AI for current detail.

Chicago: footnotes or in-text, with version specifics

Chicago has accommodated AI citation in both notes-and-bibliography and author-date variants. A notes-and-bibliography example:

1. Text generated by ChatGPT (Mar 14 version), OpenAI, 14 March 2026, chat.openai.com.

Author-date format follows similar principles to APA. Chicago's broader guidance emphasises:

• AI text is acknowledged in the body of the work as well as in the citation.
• Where AI contributed to analysis, that contribution should be described in the methodology section.
• The author bears responsibility for the accuracy of any AI-derived content.

See the Chicago Manual of Style Q&A on documentation for the current canonical guidance.

What citation styles cannot do for you

Format is the easy part. The harder part is the underlying integrity practice:

Disclosure must be specific. "AI was used to assist with research" is rarely sufficient. The expectation is that you describe what AI was used for — generating initial outlines, summarising literature, drafting code, translating sources, etc. — at the level of detail your instructor or publisher requires.

Verification is non-negotiable. Any factual claim that originated with an AI output and made it into the final work needs to have been verified against primary sources. "ChatGPT said this" is not a defence; the AI is being cited because you incorporated something it produced, which means you are responsible for whether that something is true.

Preservation is increasingly expected. Many journals and institutions now expect that an AI-assisted work can be audited — meaning you can produce the prompt, the response, the model version, and the date if questioned. This is the part most students and researchers underestimate.

Preservation: why the platform sidebar is not enough

The natural assumption is that ChatGPT or Claude's sidebar will still hold the conversation when you need it months later. In practice:

• Conversations can be deleted (intentionally or by an account change you do not control).
• Platform retention is policy-dependent and changes over time.
• Some platforms have had outages, account migrations, or feature changes that affected visible history.
• A conversation that exists in the sidebar is not the same as a conversation you can quickly find — auto-generated titles do not help you locate "the conversation from October where ChatGPT explained the regression result."

For academic citation purposes, the safe pattern is: preserve the conversation independently of the platform that hosted it.

There are three reasonable options:

1. Manual preservation. Copy the prompt and response into a methods file, an appendix, or a supplementary materials document. Date-stamped. Includes the model version.

2. Periodic export. Use the platform's data export feature periodically. Sits in your own files, searchable with standard tools.

3. Local full-text indexing. A local indexing extension like LLMnesia builds an on-device, searchable copy of every conversation. The advantage over periodic export is that retrieval is immediate — you can find "the conversation where I asked about the regression assumption" in seconds when you are writing up six months later.

For an active research project, option 3 plus periodic option 2 export is the most robust pattern.

A practical workflow for AI-assisted research

A working pattern for a typical research project where AI is used as one tool among many:

At project start: Decide which AI tools you will use and for what. Note this in your methods file from day one.

During each AI session:

• Use the AI tool normally.
• For any conversation that produced output you might incorporate, ensure it is preserved (the indexing extension does this automatically; otherwise copy or screenshot deliberately).
• Note the model version if you can — some platforms show it in the UI; otherwise the date is the next best identifier.

At write-up time:

• Search your AI history for the conversations that fed into specific claims or sections.
• Cite according to the style guide that applies.
• Add a methods or acknowledgements note describing what AI was used for at a high level.
• If your institution requires an appendix of AI prompts and outputs, assemble it from your preserved conversations.

At submission: Keep the preserved conversations for as long as the work might be audited. For published work, this can be years.

Where LLMnesia fits

LLMnesia is a Chrome extension that indexes AI conversations locally on your device — across ChatGPT, Claude, Gemini, Perplexity, and others — and gives you full-text search across them.

For academic use specifically:

• Conversations are searchable by content, not just title. Find the exact discussion about a specific paper or claim, months later, by searching a phrase from the conversation.
• Local-first. Your conversation index never leaves your device. Useful when work touches unpublished research or sensitive data.
• Cross-platform. If you used Claude for some analyses and ChatGPT for others, both are retrievable from one search.
• Persistent. Even if your platform sidebar changes or a conversation is lost on the host, the local index retains what was indexed at the time.

This does not replace deliberate preservation for the conversations you know you will cite — for those, copy into your methods file directly. It does mean you have a fighting chance of finding the conversations you did not realise would matter when you wrote them.

The bottom line

Academic citation of AI work is a moving target, but the underlying expectations are stable: disclose, document, and be able to evidence. The format your style guide requires is the easy part. The hard part is having the conversation still findable, with the prompt and response intact, when the question comes up. Treat AI conversations like any other research source: preserve them independently of the platform that hosted them, and build a retrieval system that does not depend on remembering auto-generated titles.