OpenAI and Anthropic are partnering with the Simons Foundation on a $15 million initiative to help mathematicians evaluate AI-generated research and establish standards for using artificial intelligence in mathematics.
Announced on October 15, the initiative will fund research workshops and other programmes to help mathematicians examine AI-generated results, understand how they were produced, and determine how they can contribute to existing research.
The funding follows earlier criticism of OpenAI’s mathematical research, including its claimed solution to the Navier-Stokes problem, which drew objections from mathematicians over research attribution, verification, and how the company presented its findings.
These concerns became more pressing when OpenAI faced another criticism from mathematicians over its release of hundreds of AI-generated mathematical results, some of which contained errors or were difficult to verify.
How the $15 Million Initiative Will Work
Under the partnership, the Simons Foundation, OpenAI, and Anthropic will contribute a combined $15 million in funding and resources. The foundation will manage the funding and distribute grants through an open call for proposals, while OpenAI and Anthropic will provide access to AI tools and computing resources for selected researchers.
With $5 million being OpenAI’s contribution to the initiative, the company said the funding would support workshops, conferences, and special programmes to help mathematicians examine and build on results produced by AI systems.
The initiative will also support research beyond mathematics, including work in basic science.
According to the Simons Foundation, the programme will give established institutes and research centres opportunities to share their experiences with AI tools and help shape standards for evaluating and using them.
Applications for the foundation’s Targeted Grants to Math+AI Institutes programme are open, with proposals due by December 1, 2026.
OpenAI’s Mathematical Research Faces Criticism
The announcement follows OpenAI’s October 6 release of 722 mathematical manuscripts covering work on 722 manuscripts grouped into 372 result families across mathematics and theoretical computer science. The results were generated by an internal AI model and included claimed solutions to problems that researchers had been studying for years.
However, the release drew criticism over the quality and presentation of some papers, the difficulty of understanding certain proofs, and the amount of work required to check the results. Researchers also questioned whether the release followed recommendations from an independent advisory group that had urged AI companies to prioritise human understanding when sharing mathematical discoveries.
The concerns became more pressing when OpenAI withdrew three manuscripts less than a day after publication because of a sign error that invalidated an argument in one paper and affected two related papers. The company also revised 14 other manuscripts to correct proofs, clarify statements, and address other issues.
OpenAI said it welcomed feedback from mathematicians and would continue correcting errors and improving how it presents future research.
Verification Remains a Major Challenge
One issue is that only about 42% of the 719 top-line results remaining after the withdrawals had been formalised in Lean, a programming language that allows mathematical proofs to be checked by a computer. Although Lean can help verify whether a formal proof follows mathematical rules, it does not automatically establish that the accompanying written explanation accurately represents the proof or that the result is new.
The volume of material has also placed a significant review burden on mathematicians, who must determine whether individual results are correct, understand their reasoning, and establish how they relate to existing research.
On the one hand, some researchers have welcomed the potential for AI to solve difficult problems and help mathematicians explore new approaches. On the other hand, others have criticised the way companies release large collections of results without ensuring that they are sufficiently clear, independently checked, and properly contextualised.
In an October 6 statement, AGMAI, an independent advisory group hosted by the Institute for Advanced Study, said the mathematical community still needed to assess the results and reaffirmed its recommendations for responsible AI-generated research.
Simons Foundation Wants Researchers to Lead
David Spergel, president of the Simons Foundation, said the mathematics community should lead discussions about how AI tools are evaluated, adopted and used, rather than leaving those decisions to individual companies or institutions.
The foundation has previously supported the development of Lean and other efforts connecting mathematics with AI, and its new initiative will build on that work by giving researchers funding and resources to examine AI-assisted discoveries and establish practices for using the technology.
This comes as AI companies compete to demonstrate progress on difficult scientific problems. As such, its effectiveness will depend in part on whether the additional resources help mathematicians independently assess the results and establish standards that AI developers can follow.
