Twenty-five Fields Medalists — including Terence Tao, Peter Scholze, Manjul Bhargava, and Pierre Deligne — have signed a joint declaration arguing that AI companies are damaging mathematics by treating it as a benchmark to be beaten.

The signatories don’t dispute the capability claim. They accept that LLMs have improved dramatically in recent months and can now solve major outstanding problems. Their objection is about goals, not ability: in their words, “the goals of the AI companies and the goals of the mathematical community are severely misaligned.”

The heart of the argument is that problem-solving was always a proxy, not the point:

  • Research mathematics is the pursuit of understanding — shapes, numbers, natural phenomena — and famous problems are landmarks that measure whether that understanding improved.
  • A solved problem only counts once the community digests it: the talks, the simplifications, the eventual textbook presentation a graduate student can study.
  • AI solutions get announced at speed, with no room for a proper write-up, isolation of new methods, or citation of prior work — producing what they call “severe attribution and plagiarism questions.”
  • Their warning is blunt: the mass production of true/false statements “could destroy fertile ground instead of breathing life into new ideas.”

There’s also a practical dependency being waved away. Ideas become part of mathematics through the human labour of the people who develop and integrate them. Skip that, and AI-conceived results never “become fully alive” — you get a growing pile of theorems that no one has actually absorbed.

What makes this worth reading beyond mathematics is that the signatories frame the problem as general: training in any field traditionally produced both an output and the ability to formulate new questions, and AI separates those two. That’s an alignment problem in the ordinary sense — a mismatch between what the technology optimizes for and what the work was meant to achieve — and it lands in every profession that generates a corpus worth learning from.