OpenAI Math Project Withdraws Three Papers After Sign Error

OpenAI’s math repository announces three withdrawals

OpenAI — the AI research lab — publicly announced on October 7, 2026 that it is withdrawing three manuscripts from its open‑source math repository:

• Algebraic … Weil … eightfolds – the paper that introduced a sign‑‑ sign ​‑ error that invalidated its stabilization ‑ trace cancellation ‑ argument and the construction used by two dependent papers.
• Algebraic … Kuga–Satake Correspondences for K3 Surfaces
• The rational Hodge conjecture for products of K3 surfaces

The withdrawal notice is attached to each README README README README README README.

Why the withdrawal matters

The sign error not only broke the primary result but also undermined the logical foundation of two other papers that relied on it. By withdrawing the three papers, OpenAI prevents the propagation of an invalid theorem throughout its growing corpus of AI‑generated mathematics.

Scope of the follow‑up fixes

In addition to the three withdrawals, OpenAI revised 14 other manuscripts to repair proofs, clarify hypotheses, and correct citations. The fixes fall into four thematic groups:

• Lipschitz heights and Ashkin–Teller currents (4 papers) – repaired crossing, boundary‑attachment, conditioning, and convergence arguments, and added work on the real‑Lipschitz interface proof.
• Kähler minimal model programs and abundance (6 papers) – expanded positivity and contraction arguments and clarified required input hypotheses.
• Taming and hypersymplectic deformation (2 papers) – corrected a cone‑equality claim, added a strict‑inclusion example, and removed an unnecessary cone‑comparison dependency.
• Incompressible Box Transport and Finite Computation – revised torus‑projection and common‑clock estimates.
• Exact Birch–Swinnerton‑Dyer formula – removed an obsolete citation to a now‑withdrawn supporting manuscript.

These revisions also triggered 13 additional citation updates to point to the newly‑revised companion papers.

Formalization progress

OpenAI added six new formalizations and five supporting results, raising the overall coverage of top‑line results to 300 / 719 ≈ 42 % of the repository’s claimed theorems.

Community reaction on Hacker News

The announcement sparked a lively discussion among HN commenters. Key take‑aways include:

"I think that on closer inspection, a lot of these fully AI generated proofs will fall apart. Even in Lean, you can build theories which compile but nonetheless state something different than what you actually intend." – ijustlovemath

"Without a thriving mathematical community to point out these things it would have stayed broken. With automated math that community as tao pointed out is at risk." – qoez

"3 mistakes (so far) out of ~400 is still a pretty good hit rate." – hmate9

"Proof by authority works until human mathematicians actually run the code. Back to prompt engineering." – soltanov

"All that AI compute and resources and they couldn't hold onto releasing the papers until a legitimate peer review was performed? Just had to get that PR stunt out to bump their valuation." – gt565k

Critics warned that the sheer volume of AI‑generated proofs could overwhelm traditional peer review, while others argued that a 0.75 % error rate is comparable to human‑written mathematics.

What this means for AI‑driven mathematics

• Transparency – OpenAI’s decision to attach withdrawal notices and keep archived manuscripts publicly accessible sets a precedent for openness in AI‑generated research.
• Reliance on formal verification – The incident highlights the importance of Lean or other proof assistants to catch subtle errors that may slip past informal checks.
• Community involvement – The discussion underscores that a vigilant mathematical community remains essential for validating large‑scale AI outputs.
• Future risk management – As AI systems begin to audit existing literature, the frequency of retractions may rise, emphasizing the need for systematic verification pipelines.

---

The information above is drawn directly from OpenAI’s math/history.md file (commit Oct 8 2026) and the top‑scoring Hacker News comments linked to the announcement.