OpenAI has asked an independent panel of mathematicians to review how AI-generated research results are assessed and released, a governance step that matters because the company says one internal model is producing discoveries faster than the field can absorb them. The nine-member Advisory Group on Mathematics and Artificial Intelligence can publish its advice and criticize OpenAI, but it has no decision-making power over the company and cannot set the pace of its internal research.
Evidence note: researched September 22, 2026, from OpenAI’s announcement, the advisory group’s own statement, a mathematicians’ declaration, an OpenAI technical report, and independent reporting from TechCrunch, The Atlantic, and Live Science. Toolsfine did not test the internal model, inspect unpublished proofs, or independently verify OpenAI’s claim that it resolved more than 100 long-standing problems.
The announcement at a glance
| Question | Confirmed answer |
|---|---|
| What was created? | An independent advisory group on AI’s interaction with mathematical research. |
| Who is involved? | Nine initial mathematicians, with the group hosted at the Institute for Advanced Study. |
| What can it do? | Advise on review, significance, release, attribution, professional standards, research tools, and learning. |
| What can it not do? | It cannot make company decisions or control the pace of OpenAI’s internal mathematics work. |
| Why now? | OpenAI says an internal model has produced more than 100 results that require evaluation and coordinated release. |
What OpenAI announced
On September 21, OpenAI said it was working with mathematicians who had established an independent advisory group after the company approached some prospective members. The company’s announcement says the group will help assess the significance of emerging results, advise on dissemination and academic standards, and consider how AI tools can support research and education.
The initial members include François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. OpenAI says they are unpaid, can change their own membership, may offer unsolicited advice, and are free to make that advice public.
The group’s own statement is equally important. It says the body operates independently of any AI company, will publish recommendations, and is willing to advise other companies whose models could significantly affect mathematics. Its immediate task is to recommend how OpenAI should coordinate the release of what the company describes as a large set of significant results.
The capability claim remains a claim
OpenAI says a model it began training on August 28 resolved the Navier–Stokes Millennium Prize problem and more than 100 other long-standing open problems across many areas of mathematics. Those are extraordinary vendor claims. The September 21 post does not publish the full set of problems, proofs, external reviews, or acceptance decisions, so readers should not treat “resolved” as equivalent to community verification.
TechCrunch reported that the group will assess significance and coordinate releases, while emphasizing that it cannot slow or redirect OpenAI’s work. Earlier independent reporting from The Atlantic and Live Science documented the dispute around the Navier–Stokes announcement, including questions about attribution, the speed of publication, and whether a machine-produced proof advances human understanding.
Why review is harder than checking an answer
A formal proof assistant can check whether a proof follows from stated assumptions, but that does not settle every scholarly question. Reviewers still need to determine whether the formalization represents the intended theorem, whether the result is genuinely novel, whose earlier ideas it uses, and whether the proof contains concepts that humans can understand and reuse.
OpenAI’s January report on AI as a scientific collaborator described pairing natural-language reasoning with Lean formalization to catch subtle gaps. That can raise confidence in correctness under a particular formal statement. It does not automatically establish originality, attribution, explanatory value, or acceptance by the relevant research community.
The panel’s independence has a boundary
The group’s ability to publish recommendations and criticize companies gives it more transparency than an internal committee. Its unpaid status and self-managed membership also reduce some obvious conflicts. Yet the boundary is explicit: the group has no decision-making authority, and OpenAI says it will not advise on how quickly the company advances internally.
That distinction responds only partly to the concerns raised in “A Severe Misalignment of AI in Mathematics,” a September 11 declaration signed by leading mathematicians. The declaration argues that mass-producing solutions can crowd out the slower work of explanation, attribution, training, and integrating ideas into the field. One signatory, Martin Hairer, is also an initial member of the new group.
Practical takeaways
- For researchers: preserve dated drafts, correspondence, prompts, code, and proof-assistant files so contribution and priority can be reconstructed.
- For journals and conferences: define disclosure, authorship, artifact-review, and machine-verification rules before submissions arrive at scale.
- For AI labs: separate correctness checking from novelty, attribution, significance, and responsible release decisions.
- For readers: distinguish a company’s “resolved” label from an independently reviewed and broadly accepted mathematical result.
Bottom line
The new advisory group is a concrete response to a real bottleneck: AI systems may be able to generate mathematical work faster than experts can validate, explain, attribute, and absorb it. Its public recommendations could improve the release process and create norms that extend beyond OpenAI. But advice is not control, and the headline capability claims remain unverified until the underlying results receive transparent, problem-specific scrutiny.
Sources
- OpenAI: Advisory Group on Mathematics and Artificial Intelligence — September 21, 2026.
- Advisory Group on Mathematics and Artificial Intelligence: purpose, independence, and current task — accessed September 22, 2026.
- Math and AI: A Severe Misalignment of AI in Mathematics — September 11, 2026.
- OpenAI: AI as a Scientific Collaborator — January 2026.
- TechCrunch: OpenAI forms math advisory group — September 21, 2026.
- The Atlantic: Math Can’t Go On Like This — September 16, 2026.
- Live Science: Mathematicians debate OpenAI’s Navier–Stokes claim — September 17, 2026.