The Pulse
One in Four Math Preprints Disclosed AI Use in August
Epoch AI found that 25.25% of math preprints submitted to arXiv in August 2026 acknowledged some form of AI use. The analysis also found that 5.86% credited AI with a substantial research contribution, while warning that the figures measure
AI.info Team ·
AI disclosures rise from 4% to 25% in four months
Nearly one in four math preprints submitted to arXiv in August acknowledged some form of AI use, according to an Epoch AI analysis published September 18, 2026. The share reached 25.25%, or 1,453 of the 5,754 math preprints examined for the month.
The increase follows a sharp climb from the spring. Epoch recorded AI acknowledgments in 3.89% of math preprints in April, 6.64% in May, 8.43% in June and 15.10% in July. The August figure marks more than a sixfold increase from April’s rate.
Epoch’s dataset covers every math preprint it identified on arXiv from January 2025 through August 2026. The organization’s analysis was written by Tara Abrishami.
Only 6% credited AI with major research help
General disclosure does not mean that AI produced the mathematics. Epoch divided acknowledgments into six categories: substantial research, research assistance, writing, code and computation, literature review, and formalization.
In August, 337 papers, or 5.86% of the monthly sample, described AI as making a substantial research contribution. Epoch defines that category as an acknowledgment that presents AI as essential, key, or otherwise comparable to a major research contributor. The remaining disclosures involved narrower forms of assistance or did not describe AI as central to the result.
The distinction matters because an acknowledgment can cover very different activities. A researcher may use an AI system to improve prose, search the literature, generate code, check computations, formalize an argument, or assist with a proof. Epoch places vague statements such as “the proof was found with the assistance of AI” in the research-assistance category rather than treating them as evidence of a major AI contribution.
The rise is not driven by new authors
Epoch tested whether the growing number of disclosures reflected an influx of researchers who had not previously published mathematics. It found little difference after restricting the sample to papers with at least one author who had published regularly before 2023.
For that group, the rate of AI acknowledgment stayed within 1.5 percentage points of the overall figure in each month. The result suggests that established mathematics researchers account for much of the increase in disclosed use, rather than the shift coming mainly from newly active authors.
Epoch does not interpret the result as proof that one quarter of mathematical research relies on AI. The analysis measures voluntary disclosures in preprints, so it captures changes in reporting behavior as well as changes in how researchers work.
How Epoch classified the papers
The researchers first filtered papers containing an AI-related keyword. They then used GPT-5.6 Sol to classify the acknowledgments and Claude Fable 5 as a validator. A human review of 200 acknowledgments classified as either substantial research or research assistance found 93% agreement with the model classifications.
Epoch says the method is subject to error, particularly because the line between research assistance and substantial research can be unclear. Its classifier requires language describing an AI contribution as “essential,” “key,” “crucial,” or similar before assigning the higher category.
The analysis also depends on what authors choose to disclose. A paper that used an AI system but does not mention it will not appear in the acknowledgment count, while a paper that names AI for a minor task will be included in the broader measure.
Disclosure now outpaces evidence of AI-generated mathematics
The figures show a fast change in how mathematicians describe the tools involved in producing research papers. They do not establish that AI systems are independently discovering the results reported in those papers, and Epoch does not make that claim.
What the August data establishes is narrower: 1,453 math preprints acknowledged AI use of some kind, and 337 described AI as making a substantial contribution. Those counts give publishers, reviewers, and readers a growing body of disclosures to examine, but they leave open how much each system actually contributed to the underlying mathematics.
The next question is whether disclosure practices continue to expand at the same pace and whether authors provide enough detail for readers to distinguish editing, computation, literature search, formalization, and mathematical discovery.