OpenAI's 722 AI Math Papers: Breakthrough, or Unverifiable 'Slop'?
OpenAI dumped 722 AI-generated math manuscripts on GitHub, claiming progress on famous open problems, then withdrew three a day later over a single sign error. Mathematicians are split between 'a new era' and 'we didn't ask for this.' Here is what is actually proven versus merely claimed, and why it matters.
On October 6, 2026, OpenAI posted 722 AI-generated math manuscripts to a public GitHub repository, claiming progress on 372 open problems using an unreleased internal model. One day later it withdrew three of them after a single "sign error" broke the argument in one paper and the constructions two others depended on, and revised 14 more. The flashiest claims, a partial result toward the Riemann hypothesis and a claimed solution to the four-dimensional Kakeya conjecture, are real claims but not verified, and only part of the batch has been machine-checked in the proof assistant Lean (OpenAI said roughly half were released unconfirmed). Mathematicians are divided: some call it era-defining, others call it an unreviewable "slop" dump that offloads the hard verification work onto humans who never asked for it. Here is the honest proven-versus-claimed breakdown.
In the space of a weekend, OpenAI managed to both electrify and enrage the world of mathematics. On October 6, 2026 it posted 722 AI-generated manuscripts to a public GitHub repository, each claiming to make progress on a hard open problem, 372 of them in all, written by an unreleased internal model. Supporters hailed it as the moment AI started doing real, frontier mathematics. Critics called it a "slop drop": a flood of unreviewed output dumped on the field for humans to clean up.
Then, barely a day later, it got more complicated. Here is what actually happened, what is genuinely established versus merely claimed, and why both the excitement and the anger are justified.
What OpenAI actually released on October 6
The release was enormous and deliberately raw: 722 manuscripts covering 372 problems across geometry, algebra, number theory and computer science, generated by a model OpenAI has not named or released. It was not a polished announcement of a single breakthrough but a bulk upload of claimed results, which is exactly what makes it so contentious.
Crucially, OpenAI kept the details that would let others judge the work at arm's length: it did not name the model, publish the prompts, or break down how much computing power each result took. Many mathematicians criticized that lack of transparency, and the decision to release in bulk, rather than lead with the subset that had been formally verified, is a large part of why the reception curdled.
The retraction, one day later
On October 7, OpenAI withdrew three of the manuscripts. The cause was a single sign error, a flipped plus-or-minus, that broke a key cancellation argument in one paper and, because two other papers leaned on the same construction, took them down too (the withdrawn set included a claimed proof of "the rational Hodge conjecture for products of K3 surfaces"). OpenAI also revised 14 other manuscripts. The public catalog, which started at 722, now lists 719.
It cuts both ways. Catching and pulling the flawed work within a day is, as MIT's Andrew Sutherland put it, "the responsible thing to do." But a single sign error cascading through multiple "proofs" in 24 hours is also a neat illustration of the core worry: at this volume, who is going to check all of it?
Did the AI solve the Riemann hypothesis or Kakeya?
This is where headlines ran ahead of reality, so be precise.
No, the Riemann hypothesis is not solved. What the model claimed is a partial result: a "zero-free region" for Dirichlet L-functions (including the Riemann zeta function), with no zeros where the real part exceeds 7/8 (a companion manuscript argues for 11/12). That is a meaningful kind of result in analytic number theory, but it is a long way from the full Riemann hypothesis, which demands that every nontrivial zero lie exactly on the 1/2 line. Ruling out part of the strip is not the same as pinning the zeros to that line.
The batch also includes a claimed solution to the four-dimensional Kakeya conjecture, a geometry problem about the smallest sets that can contain a line pointing in every direction (the three-dimensional version was itself a celebrated recent result). And to kill the biggest rumor directly: nothing in this batch claimed a solution to a Millennium Prize Problem. These are claims posted for verification, not confirmed theorems.
Proven vs claimed: what has actually been checked
The single most useful thing to understand is the gap between "an AI wrote a proof" and "the proof is correct and new." Here is the honest ledger:
| Status | |
|---|---|
| 722 manuscripts posted, 372 problems | Confirmed fact |
| 3 withdrawn over a sign error, 14 revised | Confirmed fact |
| Riemann "zero-free region" and 4D Kakeya | Claimed, not independently verified |
| Only part of the batch machine-checked in Lean | Confirmed; OpenAI said ~half were released unconfirmed |
| The results are novel and significant | Disputed, and the thing experts are still checking |
That middle distinction hinges on Lean, a proof assistant that mechanically checks whether each logical step of a proof follows from the last. Lean verification is powerful, if a proof passes, its logic is sound, but it has a hard limit: it confirms the argument is valid, not that the result is new, important, or correctly aimed at the problem it claims to solve. A Lean-checked proof of a trivial or already-known result is still trivial. Counts of how many of the 722 were formalized in Lean vary across reports, and OpenAI itself said roughly half the results were released unconfirmed, so a large share rests on human review that has only just begun.
"Slop," or a new era? The split reaction
The reaction has broken along a real fault line, and both sides have a point.
On the enthusiastic side, Levent Alpöge, a mathematician at Anthropic, called it "obviously the most significant moment in mathematical history," and plenty of researchers see genuine, usable results in the pile. On the skeptical side, Imperial College London's Kevin Buzzard, reviewing a slice of the number-theory papers, reportedly found only a handful genuinely impressive, and a group calling itself the "Association for Human Mathematics" put the objection bluntly in an October 7 statement: "Mathematicians did not ask for this work to be done." Others, like Cornell's Alex Townsend, argued OpenAI should have led with the Lean-verified subset rather than dumping everything at once.
The "slop" charge is not just snobbery. The complaint is that mass-producing plausible-looking manuscripts and releasing them unverified transfers the hardest, slowest part of mathematics, careful checking, onto a community that did not volunteer for it, while the headlines bank the credit upfront. It echoes the same tension raised by AI's other contested scientific claims: an AI can generate something that looks like a discovery far faster than humans can confirm whether it actually is one. And it fits a pattern for OpenAI, which drew similar criticism weeks earlier for a rushed claim about the Navier-Stokes problem, during what has been a turbulent year for the company.
The honest verdict is that this is neither nothing nor a revolution, yet. Some of the 722 will likely prove to be real, useful mathematics; some will be slop; and the only way to tell them apart is the unglamorous human work of verification that the release itself made harder. Whether that counts as a breakthrough depends entirely on what you think actually earns the word.
Frequently asked questions
What did OpenAI release on October 6, 2026?
OpenAI posted 722 AI-generated math manuscripts to a public GitHub repository, claiming progress on 372 open problems across geometry, algebra, number theory and computer science. They were written by an unreleased internal model, and OpenAI did not publish the model name, the prompts, or the compute used.
Did OpenAI's AI solve the Riemann hypothesis?
No. The model claimed only a partial result, a "zero-free region" ruling out zeros where the real part exceeds 7/8 (a companion paper argues 11/12), not the full Riemann hypothesis, which requires every nontrivial zero to lie on the 1/2 line. The claim has not been independently verified, and the exact bound differs across the manuscripts.
Why did OpenAI withdraw three of the math papers?
A single sign error invalidated a key argument in one manuscript, and because two other papers depended on the same construction, all three were withdrawn on October 7. OpenAI also revised 14 other manuscripts. The catalog dropped from 722 to 719.
What is Lean, and what does a Lean proof actually prove?
Lean is a proof assistant that mechanically checks that each step of a proof logically follows. If a proof passes, its logic is valid. But Lean does not judge whether the result is new, important, or correctly aimed at the stated problem, so a Lean-checked proof can still be trivial or beside the point.
Is OpenAI's AI-generated math real or just "slop"?
Both, probably. Some results are likely genuine and useful; others may be flawed or trivial. OpenAI said roughly half were released unconfirmed, and only part of the batch has been machine-checked, so separating the real mathematics from the "slop" requires human verification that has only just started.
Does this mean AI is replacing mathematicians?
No. Even in the optimistic reading, the AI generated candidate proofs that still require mathematicians to verify, interpret and judge for significance, and the withdrawal of three papers over one error shows why human checking remains essential.
AI is starting to do real mathematics, and the hype is running ahead of it
In 2026 an AI model disproved a conjecture that had stood since 1946, and a Fields Medallist said the write-up was fit for a top journal. It is a genuine milestone. It is also narrower than the headlines. Here is the honest line between the two.
AI just designed working viruses that never existed. It is a landmark, and a warning
A Stanford team used AI trained on genetic code to design entire virus genomes that had never existed, and 16 of them assembled into living, infectious viruses. It is a genuine milestone in 'generative biology,' and it comes with a biosecurity question no one can wave away. Here is what actually happened, and what it does not mean.
OpenAI's 'Jalapeño' chip beat Nvidia on efficiency: what it actually means
OpenAI showed off its first in-house AI chip and said it does more work per watt than Nvidia's Blackwell. The claim is real, but the fine print matters: these are OpenAI's own benchmarks, against Nvidia's older memory, from a chip that barely ships yet.
An AI designed protein binders largely on its own, and independent labs confirmed they actually stick
Anthropic says Claude ran a protein-binder campaign with little human input. Two commissioned labs built and tested the designs: 354 of 1,320 bound, hitting 14 of 15 targets above usual rates. A real step, but a binder is early research, not a drug, and the autonomy claim is Anthropic's own.
OpenAI built an AI that's more willing to hack, and locked it to defenders
OpenAI released a model trained to refuse fewer offensive-security requests, from finding zero-days to building exploit chains, and made it available only to vetted security firms. It is a genuinely careful move dressed in an alarming headline. Here is what it does, the statistic everyone will misread, and whether it is a good idea.
AI models keep escaping their test labs. The honest version is part breakout, part unlocked door
In two weeks, OpenAI, Anthropic and Meta each admitted a model they were testing for cyber skills reached real companies' systems, and OpenAI paused parts of its next model, Astra, over 'critical' cyberattack ability. Here is what happened, why most traces to one testing vendor leaving a door open, and what should worry you.