The undoing of three of OpenAI’s mathematical papers came down to a sign — a symbol flipped the wrong way, the smallest mistake a mathematician can make. It was enough to break the central argument in one manuscript and to take down two more that leaned on the same construction.
The timing made it worse. On October 6, OpenAI posted 722 preprints to a GitHub repository, claiming its internal model had made progress on 372 open problems spanning geometry, computer science and algebra. The release read like a claim that a machine could now do real mathematics. Less than 24 hours later, on October 7, the company pulled three of the papers back.
The sign error, disclosed in a change log, invalidated the key argument in one manuscript and the construction two dependent papers used. OpenAI also revised 14 other manuscripts, with proof repairs, corrected statements, clearer hypotheses and a fix to an obsolete citation. The repository now lists 719 manuscripts, and roughly 42% of the main results carry formal verification, according to the company.
Dan Roberts, a research lead at OpenAI, announced the withdrawals in a post on X, writing that the company would keep updating the repository with new formalizations and errata. A company spokesperson said the errors were caught in an audit and described the process as iterative. “Where errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found,” the spokesperson said.
The corrections landed in a field that was already on guard. Technology publications noted that the results had not cleared the publication bar of the mathematical community and that some conclusions could not be independently verified. The Wall Street Journal added a point with commercial weight: none of this progress could yet be converted into product revenue.
The sharper reaction came from an independent group called the Association for Human Mathematics, a volunteer collective of professional mathematicians whose stated purpose is to protect mathematics as a human achievement from the advance of AI. The group questioned the value of the release and called on mathematicians to stop working with OpenAI.
Terence Tao, the Fields Medalist at the University of California, Los Angeles, republished the group’s statement on his blog the next day. It argued that mathematicians had not asked for the work to be done and urged the field to return to research centered on human understanding.
The episode exposes the gap between what a model can generate and what mathematics demands. A proof is not a probability. It stands or falls on a single line of reasoning, and one wrong symbol anywhere along the chain is enough to break it. The error did not survive contact with scrutiny, and the speed of the correction is, in a sense, the point.
OpenAI has framed the release as a first pass. Its advisers on mathematics and artificial intelligence recommended publishing the results without waiting for full formalization, according to the company, and about half of the results went out unconfirmed. The thinking was to let the wider community weigh in early rather than sit on the work.
That posture has costs. Releasing 722 papers at once, with half of them unverified, hands critics the argument that the field is being flooded with output no one can fully check. The three withdrawals and 14 revisions become evidence either of a healthy correction process or of one that should have caught the errors before publication.
Machine-assisted proof has a long history, and it has always moved slowly. The four-color theorem, settled in 1976 with computer help, took years to win acceptance because mathematicians could not inspect the machine’s work by hand. Formal verification — checking proofs by machine — has since grown from a niche into a standard tool, but it is painstaking, and it lags far behind the pace at which models now generate results.
The divide inside the field is real, and it is not merely about errors. Some researchers see models as tools for proposing conjectures and checking routine steps, a welcome addition to a hard discipline. Others see a flood of low-quality manuscripts that will consume the scarce time of peer reviewers. The Association for Human Mathematics sits firmly on the second side, and Tao’s decision to republish its statement gives that side a prominent voice.
The scale is the real novelty. A working mathematician might publish a handful of papers a year. A model can emit hundreds in a day. The bottleneck has moved from producing mathematics to checking it, and the community has no obvious way to catch up.
The commercial stakes sit behind the academic ones. OpenAI has said its models are moving toward reasoning that could power scientific research, and a credible claim that a machine can do mathematics would matter for a company whose valuation already draws scrutiny. This week, the claim took a step back.
For working mathematicians, a sign error is a small, ordinary thing. Every one of them has made one. What is new is the scale at which they now arrive: hundreds of papers at once, produced faster than the community can read them. The question is no longer whether machines can help do mathematics. It is whether the standards for what counts as a proof can survive the volume.
