The Pulse
OpenAI Withdraws Three Math Manuscripts After a Sign Error
OpenAI withdrew three AI-produced mathematics manuscripts on October 7 after a sign error invalidated an argument in one paper and a construction used by two others. The company also revised 14 manuscripts and updated citations in 13 more i
AI.info Team · Written with AI ·
A correction reaches beyond one paper
OpenAI withdrew three AI-produced mathematics manuscripts on October 7 after a sign error in one paper invalidated an argument and a construction used by two dependent papers. The company’s repository history identifies the error in “Algebraicity of Weil classes on split abelian eightfolds,” then names “Algebraicity of Kuga–Satake Correspondences for K3 Surfaces” and “The rational Hodge conjecture for products of K3 surfaces” among the withdrawals. The decision puts a concrete limit on a research release that OpenAI had introduced the day before as a broad set of mathematical results from an internal model.
OpenAI’s history states: “In ‘Algebraicity of Weil classes on split abelian eightfolds’ a sign error invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers.” The withdrawn papers now carry notices describing the gap and linking to archived manuscripts, the company says. The history does not identify the two dependent papers separately from the three listed withdrawals.
“In ‘Algebraicity of Weil classes on split abelian eightfolds’ a sign error invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers.”
OpenAI, repository history
OpenAI revises 14 more manuscripts
The withdrawals were part of a larger repository update. OpenAI says it revised 14 other manuscripts, addressing proof arguments, statements, hypotheses and dependencies, as well as correcting an obsolete citation. The listed changes range from repairs to work on Lipschitz heights and Ashkin–Teller currents to expanded arguments in papers on Kähler minimal model programs and abundance.
Those updates also affected references outside the 14 revised papers: OpenAI changed 13 additional manuscripts to point to revised editions of companion work. The company’s history records six additional formalizations and five other additions covering supporting results. After the update, it puts the share of top-line results formalized at about 42%, or 300 of 719.
The release came with a verification caveat
OpenAI announced its mathematics collection on October 6, describing results produced by an internal frontier model and publishing them in a GitHub repository. Its announcement said the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking. It also said the company was consulting an independent advisory group at the Institute for Advanced Study and would continue to improve how it presents results.
The repository README makes a distinction between publication and proof checking: its results sit at different stages of verification, and not all have Lean formalizations. It also warns that some unformalized work could contain issues. The October 7 corrections show what that caveat means in practice: a flaw in one argument reached into dependent work, prompting the company to withdraw three manuscripts rather than leave them in the catalogue without qualification.
Archived papers remain part of the record
OpenAI says it will preserve earlier editions of its manuscripts and record corrections as new versions. For the three withdrawn papers, the repository says readers can find notices explaining the gap and links to the archived manuscripts. That keeps the earlier claims accessible while distinguishing them from the company’s current position.
The correction is narrower than a rejection of the full collection: OpenAI reports withdrawals of three manuscripts, revisions to 14 others and citation updates to 13 more. But the episode exposes a practical difficulty in releasing a large body of model-produced mathematics: one argument can support results beyond the paper where it first appears, so checking and correcting the surrounding dependencies matters as much as fixing the original sign.
OpenAI’s October 6 announcement and the repository README provide context on how the collection was produced and how its verification status is described.