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

A mathematician believes two of OpenAI's new machine-written manuscripts build on their own September paper without citing it. When they tried to tell someone, OpenAI Support said it could neither pass the message on nor give them an address, as the author explains in a question posted on MathOverflow that asks whether any official channel exists at all.

At a glance

• OpenAI put 722 mathematics manuscripts on GitHub on October 6, 2026, all produced by an unreleased frontier model that OpenAI has kept to itself rather than shipping.
• According to the author, Families 335 and 336 generalize two results from an arXiv paper posted on September 2, 2026, yet neither of those manuscripts mentions the earlier paper.
• Verification is uneven: the release includes Lean proofs and reasoning summaries, but how much of each result is checked varies, and OpenAI's mathematician advisers have said they want more.

If you missed the run-up: according to Lapaas Voice, the release followed weeks of criticism from mathematicians about how frontier AI companies announce mathematical breakthroughs. The complaints covered verification, attribution, transparency and commercialization. OpenAI's README on GitHub says the company tests its models on open research problems, and that it widened those tests once its existing math evaluations saturated.

The author says Families 335 and 336 extend results from a paper posted on September 2, 2026

The complaint is narrow and specific. The author posted a paper on arXiv on September 2, 2026. After OpenAI's release on October 6, they found that some of the new manuscripts, Families 335 and 336 in particular, generalize two results from that paper without mentioning it. In ordinary mathematical publishing, a missing reference like this is the kind of thing an author is asked to fix.

The author did what most people would do and wrote to OpenAI Support. Support replied that it could not forward the message to the researchers responsible for the manuscripts and could not provide a correspondence address. So the author took the question to MathOverflow. They asked whether any official or established way exists to reach the people behind the release with corrections, missing references or other scholarly concerns.

722 manuscripts in 372 families came from a model OpenAI has not released

OpenAI published 722 manuscripts on GitHub, all generated by an internal frontier model it has kept back. The counts do not quite agree. The README text gives 719 manuscripts in 372 families, while Unite.AI, AICoder and other outlets report 722, also in 372 families. Shattered.io also notes a mismatch between 372 families and 377 results, which suggests the repository was either updated or miscounted.

According to the README, nearly all results came from a single fixed procedure. The model was given approximately 4,000 problems. Each result used three hours of ChatGPT Pro thinking compute on average, and the output was grouped into families and manuscripts that had to reach an "appropriate level of significance". The README lists two exceptions: a zero-free region for the Riemann zeta function and a proof of the Hodge Conjecture for CM abelian varieties.

The manuscripts ship with Lean proofs, but verification varies from result to result

Lean is a proof assistant: you write a proof in a formal language and a program checks every step. The manuscripts come with Lean proofs and reasoning summaries, but how much is verified differs from one result to the next. According to AICoder, 235 of 372 families link to Lean and 162 papers have a formalized main result. AICoder adds that challenges can be checked with leanprover/comparator.

The README puts formalization at roughly 42% of top-line results and warns that "Some of the unformalized results could have issues". AICoder reports that none of the manuscripts has been peer reviewed. According to The Next Web, the mathematicians advising OpenAI call the release "the beginning, not the completion". Lapaas Voice reports that the advisory group said publication neither endorses the results nor certifies the papers.

The repository handles corrections through versioning instead of correspondence

According to the README, a family groups related papers: a principal result, companion arguments, consequences or alternative proofs. Each family is classified by discipline. Corrections and revisions are recorded as new versions, earlier versions stay accessible, and every manuscript directory carries a BibTeX block for citation. AICoder reports that OpenAI developed these revision and citation protocols with advice from the IAS Advisory Group on Mathematics and AI.

Think of a software changelog. Anyone can see what changed between two versions, but the file never tells you whom to email when you spot a bug. The README offers no way to contact the authors or send correspondence. It says only that OpenAI will "endeavor to fix any such issues quickly", and that the company is "exploring community-hosted repositories for these materials".

In our view, the weak point is the design. The repository has versioning and citation blocks, so it is built to take corrections. Yet the only route in it describes is a support desk that, by the author's account, cannot pass messages along. As a result, missing references will likely end up in public posts like this MathOverflow question.

Whether Families 335 and 336 change

Because the README records corrections as new versions, the visible test is whether Families 335 and 336 get a revision that cites the September 2 paper. No date has been given for such a revision, and the README still names no contact for scholarly concerns. The community-hosted repositories OpenAI says it is exploring could bring a different way to submit corrections, but no timeline has been given for them either.