Epoch AI finds AI acknowledgments in math papers jumped to 25% in four months

Mathematicians have started saying the quiet part out loud. A growing share of them now openly credit AI in their research papers.

On September 18, 2026, research group Epoch AI published an analysis of math preprints on arXiv. It found that the share of papers acknowledging AI use rose from 4% in April 2026 to 25% by August 2026.

In some corners of the discipline, the shift is even sharper. More than half of papers in three math subfields now acknowledge using AI. Differential geometry posted the most striking jump, rising from about 8% to 57%.

What Epoch actually counted

The headline number covers any acknowledgment of AI. According to Epoch, 6% of the August preprints credited AI with a substantive contribution to the research. Epoch compares that level of involvement to the role a co-author would play.

To sort these claims, Epoch built a multi-step pipeline. First, it filtered math papers on arXiv for AI-related keywords. Then it used the GPT-5.6 Sol model to categorize what kind of contribution the AI made, with Claude Fable 5 validating the results.

Each paper landed in one of six categories. These include substantial research contributions and research assistance, among others.

Human reviewers who examined the classifications agreed with them 93% of the time. The group also says it applied stricter validation to papers claiming that AI made a substantial contribution.

Not just the new kids

Acknowledgment rates were consistent among established authors who had been publishing regularly before 2023. Researchers with long track records are disclosing AI help at similar rates as everyone else.

Epoch’s analysis frames the rise as a shift in academic norms toward more transparency about AI in research workflows. The acknowledgment data alone tracks disclosure, and disclosure is what researchers choose to put in writing.

Why math is a telling test case

The subfield breakdown adds another layer. Differential geometry going from about 8% to 57% shows adoption is far from uniform.

Epoch has also made its underlying data public. The dataset includes metadata on math preprints from 2023 onward, with complete historical data going back to 1989. It is released under a Creative Commons BY license, and Epoch is directing questions to [email protected].

What this means

For the mathematics community, the immediate issue is credit. If 6% of preprints already describe co-author-level AI contributions, journals, departments and hiring committees will need clearer ways to weigh that work. Disclosure norms are evolving faster than the rules for evaluating what gets disclosed.

The measurable part is narrower and arguably more interesting: in a single four-month window, a quarter of math preprints began naming AI as part of the process. The next data points to watch are whether that 25% keeps climbing, whether other subfields follow differential geometry past the halfway mark, and whether the share of substantive contributions grows beyond 6%.