OpenAI Dumped 722 Math Papers on GitHub in One Night — and the World's Top Mathematicians Are Calling for a Boycott
Seven hundred papers landed before anyone could read them
On the evening of October 6, OpenAI opened a GitHub repository called openai/math and uploaded 722 manuscripts — organized into 372 research families, spanning algebra, number theory, theoretical computer science, mathematical logic, and topology. No press conference, no victory lap, no blog post. Just a commit. The company says each result resolves or substantially advances a previously open problem, produced by an unreleased internal “frontier” model that chewed through roughly 4,000 problems at an average of three hours of compute per solution.
Within 24 hours, a group calling itself the Association for Human Mathematics put out a statement accusing OpenAI of staging “a demonstration of power” rather than a demonstration of scholarship — and urging mathematicians to stop collaborating with the company altogether. Fields Medalists were quoted criticizing the rollout as coercive and driven by publicity instead of peer review. And Terence Tao, arguably the most prominent working mathematician today, republished the statement on his own blog and argued separately that the field needs to rethink what it even means to “solve” a problem now that solving isn’t the bottleneck anymore.
This is the second time in two months that an OpenAI math release has detonated in the middle of the mathematics community. And it’s the clearest preview yet of the era we’re entering: AI can now generate solutions faster than humans can verify them. Everything that follows — the trust, the money, the entire scientific method — is now the bottleneck.
The September prologue: Navier-Stokes and the leak accusation
To understand why the math community was primed for suspicion, rewind to September 8. OpenAI announced that its model had found a finite-time blow-up result tied to the Navier-Stokes equations, one of the seven Clay Mathematics Institute Millennium Prize Problems — each carrying a million-dollar bounty and decades of failed attempts.
About twelve hours before OpenAI’s announcement, NYU mathematician Tristan Buckmaster claimed that his and Levent Alpöge’s recent advances on Euler equations had been leaked to OpenAI days earlier and may have shaped the company’s prompts. Buckmaster later told TechCrunch that OpenAI pressured him not to credit Alpöge — who happens to work at Anthropic, OpenAI’s arch-rival. OpenAI denied pulling anything from the pair’s private work directly, while conceding it couldn’t rule out having benefited indirectly from de-identified usage data.
Three days later, more than two dozen Fields Medalists — the discipline’s highest honor, awarded to at most four people every four years — signed a declaration titled “A Severe Misalignment of AI in Mathematics.” Their complaint wasn’t that the proof was wrong. It was that a 165-page machine-generated proof, with no named authors, no attribution trail, and no time for the field to absorb it before the next release buries it, isn’t the same achievement as a result mathematicians can actually understand, check, and build on.
The attribution problem got uglier still. Reporting on an earlier August batch found that two of the ten original results leaned on existing published ideas without citing them properly — directly contradicting OpenAI’s claim that the problems had “seen no progress on the main result for at least a decade.”
What the 722-paper drop actually contains
Strip away the drama and the technical facts are these:
- 722 manuscripts, 372 families of results, covering hundreds of open problems.
- The model is internal and unnamed — reportedly begun training on August 28 and described as significantly more capable than GPT-6 Astra. No public release date. Nobody outside OpenAI can rerun the experiments.
- Not all results were formalized in Lean, the automated proof-checker. OpenAI itself says the unformalized results could contain errors — and that its own researchers hadn’t read through all the papers at release. Three papers were retracted within a day.
- The release included no prompts, no compute-time details, no cost figures — the provenance that would let anyone reproduce or audit the work.
OpenAI had consulted an advisory group at the Institute for Advanced Study before the release. On September 29, that group — the Advisory Group on Mathematics and Artificial Intelligence — published recommendations gathered from more than 600 replies from the mathematical community. Their advice was blunt: labs should deposit results in repositories they don’t control, disclose prompts and compute time and cost, formalize as much as possible, and report failure rates. On the practice of testing hard problems with private models, the group wrote that it does not endorse it.
Note what didn’t happen: none of the 722 manuscripts is confirmed to touch a Millennium Prize problem. The company’s earlier September claim — reportedly prompted by rumors that Anthropic had solved two — arrived without a named theorem, a Lean formalization, or anything close to Clay Institute certification. For scale: the only Millennium Problem ever solved, the Poincaré Conjecture, took mathematicians roughly eight years to fully verify before the Clay Institute paid out in 2010. OpenAI is asking the field to move at a very different speed.
Why mathematicians aren’t celebrating
The objection from the math community has two layers, and both matter.
The first is attribution and credit. Mathematics runs on a precise economy of credit: who proved what, when, building on whose prior work. A bulk dump from an unnamed model with no citation discipline — and credible accusations of using researchers’ unpublished work as prompt fuel — reads to mathematicians not as collaboration but as extraction. When the Association for Human Mathematics calls it “a demonstration of power,” this is what it means: the ability to bury the field’s output in a weekend, released on the company’s terms, in repositories the company controls.
The second is deeper, and it’s the one Tao is pointing at: a proof nobody can read isn’t scholarship yet. If a 165-page machine proof takes humans months to interpret — assuming it can be interpreted at all — then the field hasn’t gained understanding; it’s gained a black box with an answer attached. Mathematics, unlike model training, advances by compression into human understanding: theorems become textbooks become intuition become the next theorems. A solution that skips the understanding step is a dead end disguised as progress.
Both objections land harder because the timing looks like benchmark warfare. OpenAI says it began its large-scale assault on all Millennium Prize Problems on September 1 — days after rumors that Anthropic had cracked two of them. The math release isn’t just research; it’s a credibility weapon in the lab wars, fired on a PR timeline instead of a scholarly one.
The lesson for everyone building on AI
Here’s where this stops being a math-community drama and becomes a business problem. Every enterprise deploying AI agents right now is building on the same asymmetry: generation is cheap, verification is expensive.
- Three retractions in a day from a team that hadn’t read its own output. If the world’s most valuable AI lab ships 722 papers with a known error rate and no human review pass, what does that say about the agent-generated reports, code, and decisions flowing through your company?
- Lean formalization is the real story. The proofs machine-checked in Lean are trustworthy without human re-derivation. The rest aren’t. The takeaway: never ship AI-generated artifacts without an automated verification layer — tests for code, formal checks for proofs, reconciliation for numbers. The artifact is not the product; the artifact plus its verifier is.
- Provenance is the new citation. The IAS advisory group’s demands — disclose prompts, compute, cost, and failure rates — are a draft of what enterprise AI governance will require. If your AI workflow can’t say what prompt, model, and data produced an output, you’re one leak accusation away from Buckmaster’s situation.
- Solving isn’t the bottleneck anymore — absorption is. Tao’s point applies far beyond mathematics. In coding, the limit is no longer writing functions; it’s reviewing them. In research, it’s not generating papers; it’s deciding what’s true. In your business, it’s not producing answers; it’s knowing which answers to trust. Every team should be reallocating budget from generation to verification, right now.
What to watch next
Verification will take months — some papers are hundreds of pages of dense calculation. Watch whether independent Lean formalization efforts confirm or puncture the headline claims, whether the Association for Human Mathematics’ boycott call gains signatories, and whether Anthropic’s rumored Millennium work surfaces, turning a math dispute into a full lab-race arms spiral.
The deeper bet: the first lab to pair its AI research engine with credible, third-party verification — public formalization, open prompts, independent repositories — wins the mathematics community’s trust. That trust is the actual prize, not the million-dollar bounties.
For everyone else, the takeaway is simpler. AI has crossed the threshold where it can produce more answers than humans can check. The companies and institutions that thrive won’t be the ones with the most answers — they’ll be the ones with the best verification. Build the checking before you scale the generating.