Acknowledging AI Use
If AI helped produce part of your work, what should you say about that help, and where? The AI-Assisted Work activity showed that there is rarely a single right answer. This page offers practical advice on describing AI assistance in the cases where you’ve decided the use was appropriate.
The key word is acknowledging, not citing. In most research settings, an AI tool isn’t a source you cite for a claim, the way you’d cite a letter or an article. It’s a tool that helped you work, and what readers need is an honest account of how.
There’s no universal rule. Norms differ across disciplines, institutions, publishers, and classrooms, and they are very much in flux (Hosseini et al. 2023; Hosseini et al. 2025). What follows are suggestions on how to make these decisions responsibly.
Citing Sources, Not the Model
If a model helped you find, summarize, or interpret evidence, your citation should point to the evidence itself: the letter, the article, the archive, the dataset. AI output is not a scholarly authority on a topic, its claims need checking against sources, and once you’ve checked them, you cite the sources (as the APA’s guidance also advises).
This is why citing ChatGPT for a fact is always the wrong move: it points your reader to an answer that can’t be checked, reproduced, or trusted on its own. Some style guides do offer formats for citing AI output (APA, MLA, Chicago). They’re useful when you quote or reproduce the AI’s output, and especially when that output is itself the thing you’re discussing (for example, if you’re studying how chatbots describe Stevenson).
Acknowledging a Tool’s Role in the Work
If AI materially affected your wording, structure, code, translation, interpretation, data, or workflow, say so. The form varies with the setting (Nature 2023):
- A methods section or note, for AI used in the research process: extraction, classification, translation, analysis, code.
- An acknowledgement, for help with drafting or editing.
- A disclosure statement, where a course, journal, or funder asks for one.
- An appendix or supplementary material, for detailed prompts, settings, or evaluation results.
- Your own records, for use that doesn’t need public mention (Documenting Your Work).
Principles
- Disclosure should be proportional to impact. A minor brainstorming use may need little or no formal mention. A major analytical or drafting role should usually be disclosed more explicitly.
- Be specific. AI tools were used in preparing this article tells a reader almost nothing. Say what the tool did, on what material, and what you did with the result.
- Say what you checked. Verification is what makes AI-assisted work defensible: say how the output was checked, and what that check found.
- Neither overstate nor understate. Don’t hide a substantive contribution behind vague wording, and don’t exaggerate a trivial one.
- Policy takes precedence. When a course, publisher, journal, archive, or funder has a rule, follow it. When there’s none, use the strongest standard of transparency you’d be comfortable defending.
A Decision Framework
- What did the AI actually do: brainstorm, summarize, draft, translate, extract, code, or analyze?
- Did it provide evidence, or only language?
- Will the result stay private, or appear in a public or assessed output?
- How much of the final result depends on the AI’s wording or judgement?
- What checking was done before reuse?
- What would a reasonable reader, instructor, or reviewer need to know?
If you’ve kept a record of your use (Documenting Your Work), you’ll have the answers to most of these already.
Examples of Disclosure Statements
For analysis:
I used Gemini Notebook to generate preliminary summaries of three Stevenson letters while identifying recurring themes. All interpretive claims in the final text were checked against the letters themselves.
For brainstorming:
I used ChatGPT to suggest alternative search terms for nineteenth-century emigration in newspaper databases. Its suggestions were used only to broaden my searches, it supplied no evidence or wording in the final piece.
For data:
Named entities were extracted with a pre-trained model, and evaluated against two people’s annotations of eight letters (see Methods). An LLM (Claude) helped write and debug the extraction code, the code is available at [link].
For writing:
An early draft of the introduction was revised for clarity with the help of an AI writing assistant. The argument, sources, and final wording are my own.
You may encounter tools that claim to detect whether a text was written by AI. General-purpose AI detectors are unreliable, and widely used ones have been shown to flag writing by non-native English speakers as AI-generated far more often than writing by native speakers (Weber-Wulff et al. 2023; Liang et al. 2023). Some AI providers add watermarks to their own output, which can be detected reliably, though usually only by the provider, and less reliably after heavy editing or translation (LLM Behaviour explains how) (Kirchenbauer et al. 2024; Dathathri et al. 2024). The reliable way to show how your work was made is to say so.
Exercise
Imagine you’re publishing a short article on Stevenson’s relationships with his editors, and it includes the table of names you made and checked in steps 2 to 4. Using your log entry from step 5, write a two- or three-sentence disclosure statement for it.
Once completed, check your statement:
- Is it specific enough? Does it say which tool, and what it did?
- Does it make clear what you checked, and how?
- Does it avoid overstating or understating the AI’s role?
- Would a reader know whether the table could be trusted?
Finally, try the same for one of the scenarios in the AI-Assisted Work activity, and compare the two.
Key Takeaways
- Cite the actual sources that support your claims, not the model.
- Acknowledge AI when it materially shaped your wording, analysis, code, or data, in proportion to its impact.
- A good acknowledgement is specific: which tool, what it did, and what you checked.
- When a policy exists, it takes precedence. When it doesn’t, choose a standard you could defend.
You’ve seen what conversational AI does well, and what it can’t give you: reproducible, transparent, verifiable results. Unit 3 builds a pipeline that does. Before you start, set up Python.
If you want to read more, the following are a good place to start:
- The style guides’ own advice on citing and acknowledging AI: APA, MLA, and Chicago. Each is short, with examples. Check which one your field uses, and look for updates: all three have changed since 2023.
- Hosseini et al. (2023) and Hosseini et al. (2025): two essays, partly by the same authors, arguing first that disclosure is an ethical duty, and later that disclosing help with writing should be voluntary. Together they show how unsettled the question still is.