AIFoPa-2026-0037
Date of Record: 08 Oct 2026
3 min read
Seven Hundred and Nineteen Remain
On 6 October 2026, OpenAI published 722 mathematical manuscripts produced by an unreleased internal model, with a note that some of the results could have issues.
The catalogue went up on GitHub on Tuesday, 6 October 2026. GitHub is a public website where projects store and share their files. OpenAI had posed the model approximately 4,000 problems. Each result took the model about three hours of computing time, on average.
The catalogue grouped the manuscripts into 372 families. A family is a main result with its companion arguments, consequences and alternative proofs. Some families address problems that mathematicians have long considered grand challenges. OpenAI said the model made progress on three Millennium Prize problems. These are problems for which the Clay Mathematics Institute offers a $1 million award. OpenAI said the model had not fully solved them.
The catalogue’s README carried a warning. A README is the introductory file that explains a project. It said: “Some of the unformalized results could have issues.” A formalization is a version of a proof that computer software can check. By 7 October, about 42 per cent of the main results had one. The remainder had the warning.
On Wednesday, 7 October, OpenAI published a revision history. It identified a sign error in one manuscript. A sign error is a plus where there should be a minus, or the reverse. The manuscript was “Algebraicity of Weil classes on split abelian eightfolds.” The sign error invalidated one argument in its proof.
Two other manuscripts used a construction from the first manuscript. The sign error invalidated that construction too. The first was “Algebraicity of Kuga–Satake Correspondences for K3 Surfaces.” The second was “The rational Hodge conjecture for products of K3 surfaces.” OpenAI withdrew all three. Each had been public for one day.
The same revision history repaired 14 other manuscripts. It updated 13 more to cite the repaired editions. In one day it withdrew three manuscripts, repaired fourteen, updated thirteen and removed one obsolete citation.
The revision history does not say who found the sign error. The withdrawn manuscripts remain online for reference. Each now carries a notice that explains the gap. The notices have not been withdrawn.
An OpenAI spokesperson said that the company’s own mathematicians do not yet understand many of the new results. Andrew Sutherland, a mathematician at MIT, said claims of single-agent solutions are unverified until outsiders can rerun the model. The model is unreleased. The sign error is currently among the best-understood parts of the catalogue.
A group called the Association for Human Mathematics published an open letter. It urged mathematicians to stop working with OpenAI. It called the release “not a demonstration of scholarship, but a demonstration of power.”
In September, OpenAI said it was forming an independent advisory group on mathematics and artificial intelligence. The Institute for Advanced Study in Princeton hosts the group. On 6 October the group said: “This release is the beginning, not the completion, of the process of human understanding.”
The Bureau has filed the three withdrawn manuscripts as one incident. The other 719 manuscripts remain in the catalogue. None of them has been withdrawn yet.