OpenAI Retracts Three Mathematics Papers Within Three Days; Sign Error Causes Dependency Chain Failure

On October 7, 2026, OpenAI retracted three mathematics papers because a sign error in a foundational paper caused the entire dependency chain to fail.

Factual Reconstruction

On October 6, OpenAI published 722 mathematics manuscripts on GitHub, sourced from an unreleased in-house model. Thirty-six hours later, on October 7, the company retracted three papers, including two dependent papers, citing a sign error in “Algebraicity of Weil classes on split abelian eightfolds.” The retracted papers now include a note linking to archived versions.

On the same day, OpenAI also revised 14 other manuscripts, fixing proofs, clarifying assumptions and dependencies, and added six formal proofs, bringing the formalization rate of top-level results to 300/719, or about 42%. Previously, external researchers had been unable to reproduce the model’s outputs.

Mechanism Breakdown

The sign error directly undermined the stabilization–trace elimination argument and subsequent constructions, causing two dependent papers to fail at the same time. OpenAI also repaired other series, including topics such as Lipschitz heights and the Kähler minimal model program, indicating that once an error is found, cross-dependencies must be checked systematically.

At present, only 42% of top-level results have Lean proofs; the remaining results lack machine-verification support. This indicates that the speed of AI-generated mathematical content has far outstripped what existing verification processes can cover.

Industry Impact

The incident exposes the risk in AI mathematical research of a mismatch between claims and actual outputs. The dependency-chain failure case shows that a single foundational error can affect multiple results and increase the difficulty of external reproduction. The public GitHub history shows that the README of the retracted papers now explains the gap and provides an archive.

For researchers who rely on AI-assisted mathematical work, such retractions may reduce trust in mass-generated results while driving higher demand for formal verification tools.

Strategic Judgment

[Analysis] Based on the existing retraction and correction records, the gap between the scale of AI mathematical output and verification capacity has already become a real constraint. To improve credibility in the future, it may be necessary to increase the coverage of machine verification before publication, rather than relying only on post hoc corrections. This judgment is based on the 42% formalization rate and the fact of three retractions in this incident; it is not a definitive prediction.