OpenAI Math Papers: 3 Withdrawn After Sign Error in 722 Release

A massive drop, then a fast retreat

OpenAI's latest move into mathematical research has run into the people it most needs to win over. On October 6, the company posted 722 AI-generated mathematics manuscripts to GitHub, claiming progress on 372 families of open problems 2. The papers were produced by an unreleased internal frontier model and published under an Apache-2.0 license, which makes this the most extensive look yet at the lab's work in formal mathematics 4.

The release did not hold together for long. Within a day, OpenAI withdrew three papers after a sign error invalidated one argument and, with it, the constructions in two papers that depended on it 23. Fourteen other manuscripts were revised 2. The errors were small in number compared with the size of the dump. They still showed what many mathematicians have been warning about: once AI output arrives at volume, the work of finding mistakes falls on someone.

What OpenAI actually released

The numbers give a sense of the scale. The model was given roughly 4,000 problems during the evaluation that produced the manuscripts 4. According to the advisory group OpenAI consulted, the average result used compute equivalent to about three hours of ChatGPT Pro thinking 4. Many of the proofs were formalized in Lean, the computer proof-checking language, so a machine can verify them 4. That coverage was not complete. Two results fell outside the standard procedure, and one write-up was edited by a human 4. By one account, roughly half the papers went out without independent verification 2.

That gap matters most to critics. Cornell's Alex Townsend argued that OpenAI should have led with the Lean-verified subset rather than releasing everything at once 2. On this view, the failure was in the packaging more than the mathematics. Opening with the most defensible results would have shown real capability without asking reviewers to sort through hundreds of unchecked manuscripts.

Why mathematicians are angry

The hostility goes beyond one sign error. OpenAI is coming off a contested September 8 announcement claiming an AI system had solved a Navier-Stokes problem, a question about whether the equations describing fluid motion break down at some point 3. Leading researchers said that announcement was rushed, "leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others," and more than 8,000 researchers have endorsed that criticism 3.

The new release reopened those complaints. The Association for Human Mathematics issued a statement rejecting OpenAI's claim that the release advances the subject 1. Fields Medalist Terence Tao reshared it, which amounted to an endorsement 1. A group letter circulating online makes the point more sharply. It says mathematicians did not ask for this work, and that releasing more than 700 files at once demonstrates power rather than scholarship 2. Practitioners also object to the possibility that the models drew on their work without credit, and to the expectation that they will do the hard labor of unpacking proofs that can be nearly incomprehensible 1.

Even the advisory group OpenAI consulted kept some distance. The Advisory Group on Mathematics and Artificial Intelligence urged AI labs not to use mathematical releases as marketing tools. It also said it does not endorse testing advanced problems on models the wider research community cannot use 4. That second point is pointed criticism, since the model behind all 722 papers is not publicly available.

Credit for the correction, but not trust

Not every response was hostile. MIT's Andrew Sutherland said the quick withdrawals were "the responsible thing to do" and predicted the community would view the corrections positively 3. He was blunt that this does not settle anything. Many mathematicians, he said, are "unhappy with how OpenAI has behaved up to this point," and it will take much more to rebuild the trust the company has lost 3.

The takeaway

The accounts agree on the facts: 722 papers, three pulled, 14 revised, and a field in open revolt. They differ mainly in emphasis. Some coverage focuses on the anger 12. Others stress the scale and the formal-verification work behind the release 4, or treat the retraction as a modest point in OpenAI's favor 3.

The most reasonable reading, as analysis, is that OpenAI's technical output may well be impressive, but its deployment strategy is undercutting it. Mathematics depends on careful verification, attribution, and gradual acceptance by peers. A bulk upload from an inaccessible model, only partly verified and arriving right after a disputed headline claim, cuts against all of those norms. The sign error was a minor failure. The larger failure is a release approach that treats the research community as an audience instead of as collaborators. OpenAI will likely need to change that approach before its results are taken seriously.