Over the past year, many people have written eloquently and at length on the continuously improving power of generative AI (Large Language Models or LLMs) for research mathematics. Opinions range over the entire spectrum from negative to positive, but all seem to agree that mathematics is going through a period of enormous change, and that this will represent many challenges for us all.
The purpose of the present text is a bit more prosaic. I wanted to figure out how many mathematicians are already using generative AI for their research and how this is disclosed in their papers. As a few secondary goals, I wanted to contact a set of authors (in a very polite way) to confront them with their responsability of properly reporting their scientific methodology. My hope was that it would spark further thoughts on our future as mathematicians and the future of our field. Finally, I wanted to reach out personally to people in the hope of realising a true human contact.
For those goals, I set out a small experiment. I chose two arXiv categories, namely Classical Analysis and ODEs (math.CA) and Representation Theory (math.RT). These are the ones I tend to follow closely for my own research.
I opened all new submissions in these two categories (including cross-listings from other categories) from Monday September 14 until Friday September 18. I listed if an AI statement was present and what that statement was. This statement is typically located either at the end of the introduction, or in a separate acknowledgement or AI section.
I subsequently wrote an individual email to the authors of all these papers, asking them politely to disclose the use of AI, or to add a statement to the paper if they had not used AI. I finetuned my message a bit during this process, and the final version can be found below.
If an AI statement was already present, I congratulated the authors but also commented on it if I felt it could be improved. I initially thought of doing this for two weeks, but it turned out much more time consuming than I had anticipated. Incidentally, generative AI would have been great in automating this task, but I refrained from doing that (see my own AI statement at the end) as I thoroughly enjoyed these contacts. I looked at a total of 152 papers, with around 300 authors in total.
I summarize the results of my experiment in the following table (You can also view the full data)
| # of papers | no statements | no use of AI | other use of AI | use of AI for math | |
| math.CA | 76 | 42 | 2 | 4 | 28 |
| math.RT | 76 | 43 | 4 | 4 | 25 |
| CA+RT | 152 | 85 | 6 | 8 | 53 |
| Ratio % | 100% | 55.9% | 3.9% | 5.3% | 34.9% |
Here I divided the AI statements into four categories: no statement, an explicit statement that no AI was used, one that it was used only for language/typesetting/figures and finally a category for actual mathematical use of AI. In the accompanying file, I list all the arXiv submissions I looked at and the corresponding AI statement (if present). It is instructive to go through them to see the variety of statements.
There were only minor differences between classical analysis and representation theory. Around 1/3 papers already reported intensive AI use for the mathematics in the paper, which probably underestimates the actual use.
I received so far 33 replies to the messages I sent. They ranged from very short to very long, and sometimes I communicated back and forth several times with the authors. Several authors promised to add an AI statement in their first revision.
All my interchanges with people were pleasant, and I also received AI generated replies.
Some comments that stood out:
- Someone said AI found a link between their PhD thesis and another paper of them, which they had not observed themselves
- An arXiv editor promised to look further into the presence of AI statements
- Someone compared the influence of AI on math to what happened with the Italian school of algebraic geometry beginning of the 20th century
- I wrote at length with someone who near the end mentioned the hope that an AI editor would be more kind and empathetic in the rejection of papers than current human editors.
- Some people explicitly refuse to mention the precise LLM model they use. One referred to the lengthy statement (footnote 3) in https://arxiv.org/pdf/2608.16066
- One person asked for a videocall (still to be done).
I would like to give a few recommendations for those of us who want to write a good AI statement (a bit akin to a methodology section):
- If you did not use any AI, also state that explicitly in the paper. It sometimes takes some convincing to explain the necessity, but we are slowly moving towards a situation where AI use is becoming the standard norm (whether we like it or not).
- If you only used AI for non-mathematical purposes, such as language editing, it is in many cases not obligatory to mention it. However, from the perspective of openness, I still recommend making explicit any such use. The word proofreading can be vague. Mathematical proofreading is different than just checking the language.
- If you did use AI for mathematical purposes, be very explicit in your statement: which models did you use, for which parts of the work (which section, theorem, lemma?), in what capacity (exploratory, proving, reviewing, …). A few great examples can be found here: https://arxiv.org/pdf/2608.18024 , https://arxiv.org/pdf/2609.05669.
- Financial support by AI companies or priority access to certain models should always be explicitly disclosed.
I should add that Springer confirmed to the editorial board of the journal I am serving, that they will make AI statements obligatory for all new submissions by the end of the year.
Given the topic of this text, I will end with my own AI statement. No generative AI was used in any part of this project: all text has been written by myself, all emails were sent by myself, and all potential mistakes in my reporting are due to myself.
The email I used to contact authors:
Dear colleagues
[boldface paragraph only added if AI statement present]
Below is a standard email I am sending this week to all authors in the representation theory section of arxiv. I see that you already included an AI statement, so thank you very much for that! (add comments/suggestions on the statement)
I have seen that you have put a new preprint today on the representation theory section of arXiv. Congratulations on your new work!
I am sure you are aware of the tremendous, accelerating and complicated impact LLMs and AI have on our field (culminating so far in the news on Navier-Stokes last week).
For that reason, I plan to contact all authors in the representation theory section (fairly close to my own mathematical niche) this week with the request to add a statement on the use of generative AI in their paper.
Such a statement could be:
- No AI was used at all (e.g. see https://arxiv.org/pdf/2607.19207)
- AI was only used for typesetting and language clean-up
- AI was used in a research capacity (it is best to be very explicit, see e.g. https://arxiv.org/pdf/2609.03651, https://arxiv.org/pdf/2609.12859 , https://arxiv.org/pdf/2608.18024 or https://arxiv.org/pdf/2609.05669 for some examples)
- Funding received from AI companies, or early access to new models should be disclosed as well. Also be explicit about precisely which models you have used.
I advise to be precise in your statement if you have indeed used generative AI: for which theorems or sections did you use it, for which figures, etc.
I think it is crucial for our community of mathematicians to know how new results have been obtained, and I strongly believe such a statement will be beneficial for us all.
Feel free to forward my message to colleagues or collaborators as well.
All the best
Hendrik
Received 20 September 2026.
Add to the discussion