Epoch AI tracked explicit AI acknowledgment across arXiv math preprints and found it rose from 4% in April 2026 to 25% by August, a jump that held even when the sample was restricted to authors with publication histories predating 2023. That restriction matters: it rules out the simplest alternative explanation, that the increase is just newer, less-established researchers entering the field with different disclosure habits. Established mathematicians are now acknowledging AI assistance at meaningfully higher rates than they were five months earlier, and Epoch's tracking notes some papers crediting AI tools for research ideas rather than only for routine tasks like proof-checking or literature search. The math preprint corpus is a useful proxy because it is a field with an unusually high bar for verifiable correctness, so acknowledgment growth there is harder to dismiss as low-stakes drafting help. It does not settle how much of the underlying work AI actually contributed, since acknowledgment practices vary and self-report is not verification, but the trend line itself, quadrupling in under half a year among senior authors, is the kind of number that should reset expectations about how mathematics research is currently being done.