OpenAI published 722 AI-generated mathematics manuscripts in a single GitHub repository on October 6, 2026, and within a week the math community had answered with a public call to stop working with the company. On October 15 the Simons Foundation, Anthropic and OpenAI announced a $15 million initiative to fund the human review of AI-produced results. The episode matters beyond mathematics: it is the first large-scale test of what happens when machine-generated research arrives faster than the institutions built to verify it. The bottleneck in AI-driven science has moved from generation to absorption, and nobody has agreed on who pays for the checking.
What OpenAI actually released on October 6
The release was a git commit, not a press conference. OpenAI opened a repository at github.com/openai/math and uploaded 722 manuscripts organized into 372 families of related results, according to the company's own announcement post. The work came from an unreleased internal frontier model that had been pointed at roughly 4,000 open research problems, as The Neuron summarized. OpenAI said the output spans number theory, theoretical computer science, physics and adjacent fields, and includes solutions to, or significant progress on, hundreds of problems that were previously open.
Two details separate this from a normal preprint dump. First, the repository ships formalizations of many proofs in Lean, the proof assistant that lets a computer mechanically check each logical step rather than trusting that an argument reads plausibly. Second, OpenAI disclosed compute accounting: the average result consumed the equivalent of roughly three hours of ChatGPT Pro thinking, and the repo includes 10 abridged summaries of the model's reasoning.
OpenAI was also explicit about what the release is not. The manuscripts sit at different stages of verification, not every result has a Lean certificate, and unformalized results could still contain errors. In other words, 722 manuscripts is not 722 verified theorems. It is 722 claims, some backed by machine-checkable certificates and some not.
Why the release triggered a revolt
Two days later, on October 7, the Association for Human Mathematics (AHM) published a statement that was reposted as a guest post on Terence Tao's blog. Its two most quoted lines are blunt: "Mathematicians did not ask for this work to be done," and "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power."
The statement's central argument is procedural, not mathematical. AHM notes that the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, which OpenAI cites as its legitimacy for the release, opened its own guidance by saying frontier AI corporations should not test advanced mathematical problems on internal models. By doing exactly that, AHM argues, OpenAI showed "total disregard for the norms of scientific research." The association urged mathematicians to discontinue their work with OpenAI. It did not name a single manuscript or allege that any proof is wrong.
The grievance is cumulative, and it predates October. On September 8, OpenAI announced that an unreleased model had produced a proposed proof of the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute's Millennium Prize Problems. According to Scientific American, the run used an agent swarm of roughly 10,000 model copies burning about 88 hours of compute and 130 billion tokens. The announcement collided with work by NYU's Tristan Buckmaster and Anthropic's Levent Alpöge, who had spent close to a year on a related result. TechCrunch reported Buckmaster's allegation that OpenAI asked him to drop his collaborator's credit. OpenAI's own writeup says an internal investigation confirmed its model could not have been influenced by Buckmaster's prompts, and that the two proofs differ significantly.
By September 11, TechCrunch reported that 25 Fields medalists had signed a letter arguing that racing to announce solutions damages attribution and the human chain that transmits mathematical knowledge. OpenAI also withdrew sponsorship of a Caltech math event after criticism. October's 722-file release was the third strike, and AHM is reportedly a US-based association of about 752 members that asks its members not to work with commercial AI companies or publish AI-generated mathematical writing, according to The Indian Express.
Lean and the verification divide
The strongest counterargument to AHM is simple: if a proof is correct and mechanically checkable, its provenance matters less. A Lean certificate does not require you to trust the lab, the model or the writeup. You run the checker and it either accepts or rejects.
That argument has real force, but it has a boundary, and the boundary is exactly where this release sits. Lean verification is binary and unforgiving, and it only covers the results that were formalized. OpenAI says many proofs carry Lean certificates and more will follow, but not all of them do. A manuscript without a certificate is back to being a prose claim that a human has to read, and there are 722 of them.
This is why the phrase "AI solved hundreds of open problems" deserves scrutiny. The claim of generation is OpenAI's to make and it has the compute receipts. The claim of correctness is not established until independent reviewers, or the Lean kernel, confirm each individual result. Two things can be true at once: the model may have produced genuine advances, and the release may still be unverifiable as a bundle.
The Neuron framed the shift well. The old question was whether AI could do hard mathematics at all. The new question is how fast humans can check, understand and build on a growing pile of machine-generated research. The signal to watch is not how many manuscripts a model can emit, but how many survive serious independent scrutiny.
The $15 million Simons Foundation initiative
On October 15, the Simons Foundation announced a $15 million effort to help mathematicians navigate AI, in partnership with Anthropic and OpenAI. OpenAI is contributing $5 million, Anthropic committed $5 million of funding plus compute and Claude access, and the foundation is contributing the rest in funding and resources. The foundation will manage and award the money through an open call for proposals supporting workshops, conferences, special programs and training, and both labs will give model access and computing resources to foundation-supported researchers, extending beyond mathematics into basic science.
The quote that carries the story is from Simons Foundation president David Spergel: "This is a moment for the mathematics community to come together. The community should be driving discussions of how these tools are evaluated, adopted and used, not any single company or institution."
There is a fitting symmetry here. The Simons Foundation was an early funder of the Lean proof assistant, the same tool that makes machine-checkable AI mathematics possible in the first place. The initiative is best read not as a settlement but as triage funding: it pays for the review labor that a 722-file dump otherwise pushes onto an unpaid community.
It is also, unavoidably, a reputational repair. OpenAI had spent a month absorbing criticism over speed, credit and consent, and five million dollars aimed at the people doing the checking is a direct answer to the loudest complaint. Whether it changes the dynamics depends on whether the community treats the money as independence or as co-option. One commenter on Tao's blog, a mathematician, called it "not exactly a 'whatever it takes' moment" and suggested OpenAI could add a zero.
Why peer review cannot absorb this
Mathematics has an enormous amount of scholar-hours, but it has very few people who can evaluate a frontier result in, say, a niche of algebraic geometry or the analysis of dispersive PDEs. Peer review was built for a flow of papers measured in the thousands per year across a field, not for hundreds of manuscripts landing in one evening from a single lab.
When supply of claims outruns supply of qualified reviewers, three predictable failures follow. Review labor becomes a hidden subsidy that the institution generating the claims did not pay for. Attention becomes allocation by volume rather than by importance, because nobody can triage 722 files. And credit becomes contested, because priority races compress the months of human thinking that used to be visible in a paper's development.
None of these are unique to mathematics. The same pattern is visible in AI-generated code reviews, in security research where agents can produce vulnerability reports faster than triagers can validate them, and in any scientific field where generation is now cheap and verification is still expensive. Mathematics is simply the cleanest case, because correctness has a formal definition and because a proof assistant can settle some arguments outright.
What this means for builders and researchers
Four practical lessons transfer well beyond number theory.
Ship the checker with the claim. If you publish AI-generated research, publish the Lean proof, the test suite or the reproduction script alongside it. Verification tooling is the new currency of trust, and an unverifiable bundle invites triage rather than engagement.
Publication format is a product decision. How a lab releases a result, whether by rumor, blog post, repo or paper, determines whether experts engage or push back. A 722-file repository is legible as a demonstration of capability, and it is at least partly read that way.
Consent and attribution fights will generalize. The questions raised here about using public research in model training, about credit for near-simultaneous work and about who chose the problem set, are already live in code, art and scientific writing.
Absorption capacity is a real constraint. Teams that adopt agents for research or review should budget review time as a first-class cost, because the binding limit is no longer how fast work gets produced.
Looking ahead
The AI-versus-mathematics fight is really a preview of the governance problem for machine-generated knowledge. Generators are getting cheap and fast; checkers, human and institutional, are not. The Lean kernel solves part of this for formally verifiable results, which is why formalization keeps expanding, but the majority of science does not yet have a Lean-equivalent ground truth.
Three things to watch from here. Whether independent reviewers find errors in the unformalized manuscripts, which would sharpen the debate considerably. Whether the Advisory Group on Mathematics and Artificial Intelligence issues a further statement on the release. And whether the Simons Foundation money buys real community infrastructure, standards and shared review pipelines, or just workshops.
Terence Tao has spent October writing about exactly this transition, including pieces on the future of mathematics and on how "solving" stops being the bottleneck once machines can solve. That is the honest framing of 2026: the hard part was never generating the answers. It was always knowing which ones to believe.
Sources
- OpenAI: Sharing AI progress in mathematics (October 6, 2026, updated October 15)
- OpenAI: On the Navier-Stokes Millennium Prize Problem
- AHM statement, reposted by Terence Tao
- The Indian Express: Mathematicians did not ask for this
- Simons Foundation: $15M effort to help mathematicians navigate AI era
- Euronews: OpenAI pledges $5 million to math foundation
- The Neuron: OpenAI's AI produced 722 math manuscripts
- Scientific American: OpenAI's blockbuster Navier-Stokes proof isn't what we thought it was
- TechCrunch: OpenAI fought dirty on career-making math problem, says NYU mathematician