Writing an academic paper with AI in the room

Most academic guidance about AI is one of two extremes — total bans or breathless adoption — and neither matches what is actually happening at the keyboard. Researchers, supervisors and students are using these tools every day, often well, sometimes badly, almost never with a clear shared standard. Here is a working version of that standard, organised by the stages of a real paper.

Reading and note-taking

Using a model to summarise a paper you are about to read is fine; using it to summarise a paper you will never read and citing it is not. The distinction matters. A model summary can be wrong in subtle, important ways — it can hallucinate methodology, miss a caveat, invent a figure. A reader who treats a model paraphrase as if they have read the source has skipped the part of scholarship where you actually engage with the argument. Read the paper. Use the model to test your understanding against, not to replace it.

Outlining and structure

An outline is a place where AI is genuinely useful and largely uncontroversial. Asking a model to suggest five possible structures for an argument, or to spot a logical gap in your draft outline, is the same sort of help a thoughtful colleague provides. The thinking remains yours because you are the one choosing which structure serves the argument, and you are the one writing the argument inside it.

Drafting prose

This is the contested ground. The honest rule, in the direction major venues are already moving — Nature's January 2023 editorial and the ICMJE author recommendations both rule out listing a model as an author and require disclosure of substantive AI assistance — is this: a model may help you express your ideas, but it may not produce the ideas themselves, and any prose that survives into the final paper should be prose you can defend sentence by sentence. If a reviewer asked you why you used a particular phrase, you should know why. If a model wrote a sentence and you do not understand why it chose those words, that sentence does not yet belong in your paper.

The threshold for putting a sentence in a paper is not 'does it read well' but 'could I defend it'. AI can help you get there. It cannot do the defending.

Paraphrase without laundering

Paraphrase laundering is the practice of feeding a passage from a source into a model and treating the resulting reworded version as your own writing about the source. It is not paraphrase — it is hidden quotation, and it loses the trail that scholarship depends on. The honest version: read the source, close it, write the paraphrase from memory and from your own notes, then check it back against the original to ensure you have represented it fairly. The model is not in that loop.

Editing the final draft

Polishing prose, smoothing transitions, and improving the readability of a paper you wrote yourself is editing — the same thing a copyeditor at a journal would do for you. This is where AI tools, used carefully, save real time without changing the scholarship. Two safeguards make this defensible: do not let the tool change claims, citations or numbers; and read every edited sentence to ensure you can still defend it.

Disclosure and your supervisor

Journal and university policies are moving fast. The safe rule today is to disclose materially — to disclose AI use that shaped the substance or the prose, briefly, in your methods or acknowledgements, in the form your venue prefers. A short, honest disclosure protects you. A buried one, or none at all when one is required, can end a paper and a career.

Used this way, AI is closer to a calculator than to a co-author. It speeds up real work without doing the work itself, and that distinction — between assistance and authorship — is the one a serious researcher should be able to make for every paragraph they submit.

❦