Mathematicians are adopting AI tools faster than even the researchers tracking the field expected

Epoch AI, a research organization that monitors AI progress, reports that in three of the 18 mathematical subfields it tracks, more than half of the papers uploaded to arXiv by established researchers now acknowledge AI use. The finding comes from analyzing acknowledgment sections, not from automated detection of model-generated text.

Differential geometry — the study of geometric structures on smooth manifolds — shows the sharpest jump. The share of papers reporting AI use climbed from roughly 8% in July to roughly 57% in September, a span of two months. The other two subfields that crossed the 50% threshold were not named in the release.

What the metric actually captures is an explicit author statement, not the technical presence of AI-generated code or prose. Researchers may credit a large language model for drafting, translation, bibliography formatting, or writing code snippets for numerical checks, without the model contributing to the mathematical idea itself. Epoch AI has not published its full counting methodology, nor does it distinguish among different types of use.

A shift from single digits to a majority in a matter of weeks is unusual even by the standards of new software adoption in science. The figure may reflect a change in reporting norms — more authors choosing to disclose use that already existed — alongside genuine adoption. Without baseline data from before July, the two effects cannot be separated.

The picture is partial. Epoch AI monitors only 18 subfields, and arXiv does not cover all mathematical publications. The data covers only "established researchers," a category the organization does not define, and excludes students and postdocs, who may be the true early adopters. Even with those caveats, the trend is clear: AI tools are entering the daily practice of mathematics faster than any previous technology.