OpenAI, Anthropic and Simons Back $15M AI Maths Initiative

When an AI lab claims its model has produced hundreds of mathematical results, the hard part is no longer generating them but checking them. That tension came to a head this week: OpenAI pledged $5 million to a Simons Foundation initiative meant to help mathematicians examine and build on AI-generated work amid criticism from parts of the mathematics community over how that work was released. The initiative, backed by Anthropic as well, totals $15 million and is one of the first organised attempts to build shared norms for evaluating machine-produced research.

What was announced

The Simons Foundation announced on 15 October 2026 that it, Anthropic and OpenAI will commit a combined $15 million in funding and resources. The foundation will manage and award the money, including through an open call for proposals. According to its announcement, the funds will support workshops and initiatives that help researchers assess recent AI-assisted results, learn and use new tools, share perspectives, and develop shared standards for AI use. Anthropic and OpenAI will also provide AI tools and computing resources to researchers the foundation supports, a provision the foundation says extends beyond mathematics to basic science.

The split between the three parties is not specified in the foundation's text. Euronews reports that OpenAI's donation is $5 million, but the foundation's own announcement does not break down the shares or say how much Anthropic is contributing. The foundation's president, David Spergel, is quoted saying the community "should be driving discussions of how these tools are evaluated, adopted and used."

The controversy behind it

The pledge arrived amid criticism. Euronews, in a report dated 16 October, says that in September OpenAI stated that an unreleased AI model had solved a major problem involving the Navier-Stokes equations, which describe fluid motion and have puzzled mathematicians for nearly a century. On 6 October, according to the same report, OpenAI said the model had published more than 700 manuscripts and produced almost 400 other mathematical results. These are OpenAI's claims as relayed by Euronews; the company's own post was not accessible for this article, and the correctness of the results has not been independently confirmed in the sources reviewed.

The same report says two human mathematicians told TechCrunch they had been close to a similar result and questioned whether OpenAI's work drew on theirs, and that OpenAI says it never saw their work. One of them, Levent Alpöge, works at Anthropic and was collaborating in a private capacity, Euronews notes. The Association for Human Mathematics criticised OpenAI and alleged plagiarism, saying "Mathematicians did not ask for this work to be done" and accusing the company of a "total disregard for the norms of scientific research." Those are the association's allegations; OpenAI has disputed the claim that it saw the other group's work.

Why verification is the bottleneck

The core problem is scale. A single proof can take specialists weeks or months to referee, and hundreds of machine-produced manuscripts and results arriving within weeks exceed what the usual peer-review pipeline can absorb. The Simons Foundation's announcement points to one existing tool: the foundation funded development of the Lean proof assistant, which it says has transformed how mathematicians work and made AI-assisted proofs possible. In Lean, a proof is checked mechanically, so a formalised result does not depend on a human reading every step. That does not remove the need for humans, however. A formal proof only establishes that a statement follows from its stated assumptions. People still have to judge whether the statement formalised is the one that matters, whether the result is new, and how it connects to prior work.

Credit and priority are a separate, social problem. The dispute over whether the model's output overlapped with human efforts illustrates how hard it is to establish novelty when a system is trained on much of the existing literature. Questions about attribution, disclosure of training data and whether results should be released to the community before verification are not purely technical, and they are exactly what the foundation says workshops should address through "shared standards for AI use."

How to read the initiative

There are two reasonable readings. One is constructive: the money goes to a neutral foundation, is allocated through an open call, and funds community convenings in which mathematicians, not the labs, set evaluation norms. The foundation's framing, that the community should drive these discussions, supports that view. The other is more sceptical: a donation by a company whose release practices drew criticism, announced after that criticism, does not by itself resolve the underlying complaints. Euronews frames it as an attempt to address a backlash. Both can be true, and the practical test is whether the funded work produces standards that labs actually follow.

Anthropic's participation also complicates the picture of a simple lab-versus-community dispute. One of the mathematicians involved in the priority question works at Anthropic, and the company is now a co-funder of the initiative, which suggests the issue cuts across labs as well as between labs and academia.

What to watch next

Several signals will show whether this becomes a real change in research practice. The first is the open call for proposals: its terms, timetable and review process will indicate how much influence the donors have over what gets funded. The second is whether the results OpenAI described are independently checked, ideally with formal proofs, and what the outcome is for the specific claims that have been disputed. The third is whether journals, preprint servers and conferences adopt disclosure requirements for AI-generated mathematics, such as stating which parts were machine-produced and how they were verified. Finally, the dispute over priority claims will test whether the field can settle attribution when the author is a model and the training corpus is the literature itself.

For now, the facts that can be stated with confidence are limited: a $15 million, three-party initiative exists; its stated purposes are assessment, tool training and standards; and a public disagreement over OpenAI's claims and conduct preceded it. Whether the underlying mathematical results hold up remains the open question, and the answer will come from the slow work of verification the initiative is meant to support.