TacoYaci Math
Keeping pace without AI? The AI-free caucus on arXiv
The Association for Human Mathematics lists 809 mathematicians, and 532 of them, the AI-free caucus, state that they do not use AI models in their research. Here I compare their arXiv output with that of comparable mathematicians: all other authors of arXiv mathematics papers, reweighted to the same fields, career stage and earlier output. The analysis plan was fixed before any result was computed. Hover, tap or use the arrow keys on a chart to read the numbers.
Before and after: the caucus kept pace
Papers per author and year, relative to October 2017 to September 2022. Periods run from October to September and are named after the year in which they end. Shaded: the two periods that are compared.
Show the numbers
Month by month: a gap opens in the summer of 2026
Papers per author, three-month average, relative to 2017 to 2022. Triangles mark model releases of OpenAI, Anthropic and Google. Shaded: July to September 2026.
About the peak of the caucus at the end of 2025: a few members had a very productive quarter at the same time, mostly with coauthors and in quite different fields. None of these papers mentions an AI tool, and the peak starts in October 2025, a month before Gemini 3, so the timing does not fit a model release (exploratory check).
Show the numbers
All arXiv mathematics: twice as many papers in a year
Every arXiv paper with a mathematics category, cross-lists included. Not a group comparison.
Show the numbers
Traces of AI, by group
Averages over the authors with a paper in the period. Comparable mathematicians reweighted to the caucus.
Show the numbers
Sudden jumps: no excess in the caucus
Share of authors whose output in April to September at least doubled (or tripled) against their own half-year average of 2017 to 2022. Light bars: average of 2019 to 2025. Dark bars: 2026.
Show the numbers
The whole story in one image
For sharing. Click the image for the full size.
Read with care
- Only groups are compared. Nothing here says anything about a single person, and a jump in one author's output has many ordinary reasons.
- Membership is a choice, so a difference between the groups is not an effect of avoiding AI. The other AHM members, for example, grew faster than comparable mathematicians.
- The summer of 2026 is three months of data, and its interval only just excludes no difference.
- Authors are recognized by their names, which merges people with the same name. Details, all further questions and the limits are in the paper.