OpenAI withdrew three AI-generated mathematics manuscripts from its 722-paper release after identifying a sign error in a central proof argument. The removals were recorded a day after the collection launched.
The error appeared in “Algebraicity of Weil classes on split abelian eightfolds”. According to OpenAI's version log, the mistake invalidated an argument based on cancellation in a stabilization trace.
That failure also affected two papers that depended on the argument. One of them was “The rational Hodge conjecture for products of K3 surfaces”. The source material does not identify the third withdrawn manuscript.
What changed in the release
Following the withdrawals, the catalogue contained 719 manuscripts. OpenAI's version log also recorded 14 additional proof repairs, separate from the three manuscripts removed after the sign error was identified.
The episode shows how an error in a mathematical proof can spread beyond the paper where it first appears. Here, the disputed argument was used by dependent work, meaning the problem reached multiple manuscripts rather than remaining confined to one result.
The available information does not establish whether the withdrawn papers will return after revision. It records the removals, the affected proof, the dependent manuscripts and the additional repairs, but gives no further timetable or explanation of the next review stage.
Conclusion
OpenAI's mathematics release was reduced to 719 manuscripts after a sign error undermined a stabilization-trace cancellation argument and affected two dependent papers. The version log also lists 14 further proof repairs.
Frequently Asked Questions
Q. How many manuscripts were in OpenAI's original mathematics release?
The release contained 722 manuscripts.
Q. How many manuscripts did OpenAI withdraw?
OpenAI withdrew three manuscripts.
Q. Which paper contained the sign error?
The error was identified in “Algebraicity of Weil classes on split abelian eightfolds”.
Q. What kind of argument was invalidated?
The sign error invalidated a stabilization-trace cancellation argument.
Q. How many manuscripts remained after the withdrawals?
The catalogue contained 719 manuscripts after the withdrawals.
Q. Were other proof problems recorded?
Yes. The version log recorded 14 further proof repairs.
