SERIES · AI is making papers cheap to produce but they remain costly to referee. Proofs and Prompts is publishing a series of posts from 26 October on the future of publishing, papers and credit. Questions such as how to disclose AI use, who bears the cost of checking, what authorship means and what papers of the future might look like will be addressed. This post is part of the series. Each is written independently and reflects only its authors’ views. All posts in the series
AI disclosure. This essay was composed on the Notes app of the author’s iPhone. This app provides AI assistance with spelling, grammar, and auto-completion of words being typed.
Most discussions about LLMs in mathematics seem to focus on their (rapidly advancing) proof discovery abilities and the effects these abilities may have on professional mathematics. I write here about a different use of these technologies, to make the following point: we have a professional duty to use LLMs to proofread our work. I use “proofread” to mean both copy editing of no mathematical substance, as well as the detection and correction of mathematical errors and gaps in arguments. By “our work,” I mean our research papers, though the arguments below extend with straightforward modifications to teaching materials.
First, a fundamental community norm is that we should make all reasonable efforts to ensure that our proofs are correct. Modern LLMs are tools that, when used appropriately, can provide this assurance with high confidence and little effort.
Second, incorrect or inadequately justified proofs exact a substantial human cost. I have refereed many papers in my career; the vast majority had at least some minor mathematical issues, and a not insignificant minority had issues that required new contributions by the author(s) to patch. Identifying these issues and writing a detailed referee report to alert the author(s) takes many hours, especially for long and technical papers.
Before LLMs, I executed my referee responsibilities with pride, knowing that authors were (typically) putting forth the best work they could, and that operating on the frontier of mathematical knowledge would inevitably involve some mistakes. Now, however, when I see clear mistakes in papers I am reviewing, I just feel disrespected. The author(s) had a tool they could have used to eliminate the mistakes and did not, wasting my time. (Certainly, LLMs are not perfect, but it is easy to tell when a cursory check has not been done.)
Recent events in my personal life have emphasized for me the value and precariousness of human life, and how death and life-limiting illness can strike unpredictably. Submitting work that has not gone through an LLM check to a journal may cost a conscientious referee many hours of their time. To ask someone to spend their most precious resource correcting errors that could have been caught with the push of a button is both rude and professionally negligent.
Some may argue that abstention from LLM checking is justified on ethical grounds. If this objection rests on the behavior of frontier labs, one can simply use open-weight models. If the objection rests on resource use, a simple calculation shows that the resource use from a few minutes of LLM computation pales in comparison to other activities we regularly undertake to advance mathematics, such as conference travel. I am sure there will be other creative objections, but it is hard for me to see how they can outweigh the cost refereeing erroneous mathematics has on human life (especially in combination with the cost to all other readers who might be interested in the result).
I also believe we have a duty to use LLMs for literature searches, to ensure correct attribution, and to make sure our arguments are as simple and straightforward as possible. The reasons are similar.
Received 30 September 2026.
Add to the discussion