Something notable happened on September 21 that got less attention than it deserves. A group of nine elite mathematicians announced the Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton’s Institute for Advanced Study, with a specific and immediate mandate: advise OpenAI on how to responsibly release “a large number of significant results in mathematics that they report have been produced by their internal model.”
Read that sentence carefully. Not “help evaluate whether AI can do math.” Not “develop benchmarks for mathematical reasoning.” OpenAI already has a pile of results and doesn’t know how to release them. That’s a different situation.
The board’s membership is not decorative. Timothy Gowers and Martin Hairer are Fields Medalists. Edward Witten is arguably the most influential living theoretical physicist. Ravi Vakil runs Stanford’s algebraic geometry program. Melanie Matchett Wood is one of the stronger analytic number theorists of her generation. These are not people who sign their names to institutional statements lightly.
The practical questions the group is presumably grappling with are genuinely hard ones. Mathematics is one of the few domains where AI output can be rigorously verified — a proof is correct or it isn’t, and there are formal proof assistants to check — but “correct” doesn’t resolve the harder questions: What does authorship mean when a model produces a theorem? How should journals handle AI-generated proofs that no human has produced independently? If a result is significant, does it matter that the discoverer cannot explain its own reasoning? What credit, if any, accrues to the AI’s trainers?
This is also the next chapter in a story this site has been following. In September 2026 alone, a declaration signed by 25 Fields Medalists warned that AI labs optimizing for benchmark performance were damaging mathematics by bypassing the understanding chain that makes proofs educationally useful. Po-Shen Loh’s guest post on Tao’s blog asked the question in the title directly. Now there is a formal body addressing the institutional mechanics.
The advisory group is structurally designed to stay credible. Members accept no payment, recommendations are published publicly, and the group explicitly welcomes input from the broader mathematical community. That last part matters: if the group were issuing guidance in private, it would look like cover for a coordinated announcement. Doing it in public makes it something closer to a community process.
Mathematics is an interesting test case for AI-produced knowledge precisely because the verification bar is so much lower than in empirical sciences. You don’t need to replicate an experiment or wait for a clinical trial — you need to check the proof. If AI is producing results that a board of this caliber finds worth creating a formal body to handle, the question is not whether AI can do mathematics. It’s what the discipline looks like once it can, and whether the institutions that transmit mathematical knowledge can adapt quickly enough to stay coherent.
That question has no clean answer yet. At least now there is an organized group asking it seriously.