OpenAI Withdraws Three of Its 722 AI Math Papers Over a Sign Error
One day after releasing 722 AI-generated math manuscripts, OpenAI withdrew three after a sign error in a Weil-classes proof collapsed the argument. The withdrawal came as mathematicians raised concerns about vetting, and as the New York Times reported the internal model had attempted around 4,000 research problems before the release.
On this page
A sign error, one day later
OpenAI withdrew three of the 722 AI-generated mathematics manuscripts it had published on 6 October, removing them from the version log on 7 October after a sign error in a proof about Weil classes, in a paper titled Algebraicity of..., caused the argument to collapse. The retraction arrived exactly one day into an experiment this blog covered at release: 372 result families claiming solutions or substantial progress on major open problems, attributed to a single prompt given to a single internal model.
The mechanics of the failure are worth stating plainly. A sign error is the kind of defect that Lean-checked formalizations are supposed to catch, which raises the question of whether the three withdrawn manuscripts were among those formally verified or among the majority that were not; the original release acknowledged that many results had not yet been formalized at all.
The scale question underneath
The New York Times' report on the release added a number that changes how the withdrawal reads: the internal model had attempted around 4,000 research problems during testing before OpenAI published the 372 families that survived. A yield of roughly one in ten, with the failures only visible to the company that ran the attempts, is exactly why the vetting dispute matters more than the individual retraction. The Guardian reported leaders in the mathematics community expressing concern that OpenAI was not doing due diligence to vet results, and that the models remained inaccessible to the broader mathematical community that would have to check them.
To its credit, the correction process worked as the release design promised: the version log is public, the withdrawal is documented, and nothing had to be pried out through peer review. A private lab claiming results in a private review queue could have quietly dropped three papers and said nothing.
What vetting was supposed to look like
The independent advisory group OpenAI formed after the Navier-Stokes controversy recommended releasing the model, the exact prompts, and per-result compute alongside the papers, on the theory that verifiability requires more than a repository of conclusions. So far the company has disclosed average compute statistics only and says it is not bound by the guidelines. The three withdrawals are the first empirical data point for the skeptics' position: a sign error survived whatever internal checks existed, in a field where a single flipped symbol can invalidate everything downstream of it.
The practical takeaway for anyone consuming AI-generated research is unchanged from the day of the release: Lean-verified claims are the ones to lean on, everything else is a draft, and the rate at which a lab publishes should be weighed against the rate at which it retracts. The version log now provides both numbers, which is the small good news inside the withdrawal.