A group of mathematicians, with the International Mathematical Union behind them, issued the Leiden Declaration on 2 June warning that AI threatens the field's norms around proof, attribution and who sets the research agenda.
The timing is the point. FrontierMath, Epoch AI's benchmark of original unpublished problems that take researchers hours or days to crack, sat near zero when it launched in late 2024; as of late May 2026 GPT-5.5 Pro leads it at 52.4%. The declaration's worry is not that machines prove theorems but that they generate plausible, hard-to-check proofs faster than humans can verify them, eroding trust in the literature.
It is a rare case of a discipline trying to write its own rules before the tools rewrite them, though the declaration carries no enforcement, only the IMU's weight behind it.