Skip to content

OpenAI

OpenAI’s 722 math papers come with no one to write to

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: online-tech-tips.com

The author of a paper posted to arXiv on September 2, 2026 says two entries in OpenAI’s October 6 math release, Families 335 and 336, generalize two of that paper’s results without citing it. The author says they then hit a dead end: OpenAI Support told them it could neither forward the message to the responsible researchers nor give them a correspondence address.

The author laid this out in a question on MathOverflow. It asks whether there is “an official or established way to contact the authors or researchers responsible for these manuscripts to report corrections, missing references, or other scholarly concerns.”

The release put 722 manuscripts on GitHub, all from an unreleased frontier model that OpenAI kept to itself. The repository’s readme says the model was posed about 4,000 problems, and each result used three hours of ChatGPT Pro thinking compute on average. A “family” in the catalogue groups related papers: a principal result, companion arguments, consequences or alternative proofs. So the two families the author names may cover more than two manuscripts.

The readme does set out how errors are handled. OpenAI says it will keep the public release history, record corrections and revisions as new versions and leave earlier versions accessible. Each manuscript directory carries a BibTeX block so others can cite it, and OpenAI says it is also exploring community-hosted repositories for the material. About 42% of top-line results have Lean formalizations, meaning a proof assistant has checked them mechanically. On the rest, the readme says: “Some of the unformalized results could have issues. We will endeavor to fix any such issues quickly.”

That process ran a day after the release. On October 7, a single sign error pulled three manuscripts: it broke the argument in one paper and the shared construction two others depended on. NYU Shanghai’s RITS reports that an internal audit found the errors, that the catalogue now lists 719 manuscripts, and that fourteen more were revised. Every fix it describes came from inside OpenAI.

OpenAI has been here before. TechCrunch recalls that then-VP Kevin Weil posted on X that “GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others”. According to TechCrunch, GPT-5 had only found solutions that already existed in the literature, and Weil took the post down after Yann LeCun and Demis Hassabis mocked it.

Outside mathematicians expected problems with this release. MIT’s Andrew Sutherland said “we should expect some of the proofs to contain mistakes, possibly serious ones,” RITS reports. Cornell’s Alex Townsend said he suspects more errors will surface and that OpenAI “should have announced the manuscripts that were lean verified first.” The mathematicians advising OpenAI called the release “the beginning, not the completion,” The Next Web reported.

The same RITS piece says the release has sharpened a dispute among mathematicians over how AI-generated results should be published. It quotes NYU’s Roland Bauerschmidt on part of the output: “If I had received these by e-mail from a nobody, I probably would have deleted it, it’s so badly written.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.