OpenAI Creates Independent Math Advisory Group After Claiming 100 Solved Problems

OpenAI announced on September 21, 2026 an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, New Jersey.

The group is intended to give mathematicians a channel into OpenAI's math research. OpenAI lists the notice in its newsroom as "Advisory group on mathematics and artificial intelligence" and published the full text at its index site OpenAI.

The announcement came after OpenAI published a claimed solution to the Navier-Stokes Millennium Prize problem, a famously hard question about fluid motion that has resisted proof for decades. OpenAI says the same internal model behind that claim has solved more than 100 other open problems across most areas of mathematics TechCrunch.

OpenAI gave detail on method. It said it used as many as 10,000 AI agents, or separate AI workers operating together The New York Times, and that the system reached the Navier-Stokes result in 88 hours The Guardian. On August 1, the company had published "Ten advances in mathematics and theoretical computer science," describing new results on long-standing open problems, including work in geometry.

The advisory group will assess how important new results are and coordinate how they are released. Nine mathematicians were named as starting members.

The terms are brief. Members will not be paid. They may give advice without being asked, speak publicly about their views, and control who joins. One topic is explicitly excluded. The group will not advise OpenAI on how fast to push ahead with its internal work in mathematics.

That limit is notable because many mathematicians have objected. Twenty-five winners of the Fields Medal, a leading award in mathematics, signed an open letter saying AI labs are threatening their intellectual work. Camillo de Lellis of the Institute for Advanced Study both signed that letter and joined the advisory group as an initial member.

The broader context here is checking work at machine speed. A single model run that claims to settle dozens of open problems in different specialties creates a bottleneck. Journals, outside referees, proof-checking software, and informal networks of experts all work on human timescales. In that setting, deciding the order, completeness, and readiness of releases is practical infrastructure, not public relations. Those choices will affect how quickly others can check, reproduce, or build on the work.

Looking at lab governance, the exclusion tells as much as the mandate. OpenAI has given outsiders a role in sorting and communicating results while keeping control of research pace. Readers who follow security practice will recognize the pattern. It is similar to disclosure programs and model evaluations where outside experts describe behavior but do not set the speed of training or deployment. Independence here is procedural, based on unpaid status, the right to dissent in public, and self-chosen membership, rather than veto power.

In my view, the setup deserves narrow trust. Unpaid advisers who govern themselves and can speak publicly are in a stronger position to stay credible than an internal review board. Still, asking nine initial members to judge importance across most of mathematics is its own scaling problem. Judgments about importance in mathematics depend on deep specialty knowledge. A small cross-field panel can route and prioritize, but detailed checking will still depend on wider research communities.

Looking ahead, what this could enable, if the results stand up to inspection, is a faster loop between machine-made conjectures and proofs and human understanding. I have watched my own children move from calculators to search engines to language models for schoolwork, each time amid warnings that the tool would undercut learning. Each time the lasting change was different. The tool made routine work cheaper and moved human effort toward judgment. Mathematics may now face that change in its clearest form, where routine derivation costs less and taste, choice of problems, and verification matter more.