The OpenAI math advisory group announced on September 21 is meant to give mathematicians a say in how the company reviews and presents AI-generated results. Its remit has a conspicuous boundary: OpenAI says the group will not advise on the pace of its internal mathematics research.
The Advisory Group on Mathematics and Artificial Intelligence, or AGMAI, is an independent body of nine mathematicians, according to OpenAI and a September 21 post announcing the group. OpenAI says the panel may assess the significance of emerging work, advise on review and academic standards, and help coordinate how results are communicated and disseminated. It may also advise on how the company’s tools could support mathematical research and learning.
That gives the group a role in review, communication and release coordination, while leaving OpenAI’s internal research pace outside its stated authority. OpenAI says members can offer unsolicited advice, publish it, and publicly criticize the company. It also says they will not be paid by OpenAI and can determine the group’s membership themselves.
What can the OpenAI math advisory group do?
AGMAI can advise OpenAI and potentially other frontier AI labs on how mathematical claims are checked, explained and released. It has no stated role in telling OpenAI how quickly to advance its internal mathematics work. The initial members are François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten and Melanie Matchett Wood.
The distinction is relevant because OpenAI has made unusually broad claims about an unreleased internal model. The company says it began training that model on August 28 and that it has resolved more than 100 long-standing open problems across most areas of mathematics. That is an OpenAI claim, not a result independently verified in the available reporting.
Questions about AGMAI’s independence surfaced immediately. The Verge reported that OpenAI’s rollout led some mathematicians to assume the group was company-appointed. AGMAI’s website said OpenAI had approached some eventual members about creating an advisory board before they opted to establish an independent group and invite others. It remains unclear which members were initially approached.
Hairer, an AGMAI member who teaches at Imperial College London and EPFL, told The Verge that the group receives no financial, technical or other support from OpenAI. He said members do not work for the company and have signed no restrictions beyond standard confidentiality requirements needed for advance access to research. Hairer also said the group was open to talks with other frontier labs.
Why are mathematicians questioning OpenAI?
The advisory group does not settle a separate dispute over training data and credit. Mathematician Andreas Thom questioned whether his and colleagues’ ChatGPT interactions could have influenced an OpenAI result on non-sofic groups, The Verge reported. Thom said the company’s response did not establish whether those interactions entered data used to improve models. The report does not establish that OpenAI used his work or chat data.
In an earlier dispute over a Navier-Stokes result, OpenAI said its researchers and agents had not accessed specific user data to solve the problem. It also said it could not rule out that de-identified data derived from product use helped improve its models. AGMAI’s review-and-communication role leaves those data-provenance questions unresolved.
This story draws on original reporting from The Verge.
