OpenAI withdraws 3 of its 722 AI-written math papers over a sign error

OpenAI’s library of AI-generated mathematics isn’t 722 manuscripts any more. On October 7 the company withdrew three papers over a sign error, revised 14 others, and the repository’s own README now puts the catalogue at 719. The published coverage all says 722, which is what the README said when the repo went live the day before.

The retraction starts in a paper titled “Algebraicity of Weil classes on split abelian eightfolds”. OpenAI’s history file says a sign error there “invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers”. Those two came out with it: “Algebraicity of Kuga-Satake Correspondences for K3 Surfaces” and “The rational Hodge conjecture for products of K3 surfaces”. None of the three appears in the manuscript map now.

Repository figure	At launch, October 6	After the revision, October 7
Manuscripts in catalogue	722	719
Result families	372	372
Top-line results with Lean proofs	“Many, but not all”	300 of 719, about 42%
Manuscripts withdrawn	0	3
Manuscripts repaired or re-cited	0	14 proof repairs, 13 citation updates

The formalization line moved the other way, and it’s the number worth keeping. At launch the README said only that “Many, but not all, of the manuscripts have been formalized.” The revised text swaps that for a count: 300 of 719 top-line results, which the company rounds to 42%. Six of those Lean formalizations were added in the October 7 update.

This release is the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge.

Advisory Group on Mathematics and Artificial Intelligence, October 6 statement

That group, AGMAI, is nine mathematicians hosted at the Institute for Advanced Study. Its site says the members came together after OpenAI approached some of them about an external advisory board, and they chose to form an independent group instead. Its September 29 guidelines were informed by more than six hundred survey responses. The same statement says its advisory role “should not be interpreted as a judgment of the impact of these results or an endorsement of the process by which OpenAI obtained them”.

Those guidelines asked labs to publish through established academic channels where possible, as The Verge reported, and a GitHub repo isn’t one of them. If you want the map of where mathematics actually gets refereed, we laid it out in our guide to following AI research. The venue matters here because the repository is also the correction mechanism.

The flagship result is family 003, the quasi-Riemann hypothesis, and the primary is narrower than the coverage suggests. The manuscript map says it proves every Dirichlet L-function, including the zeta function, has no zeros where the real part exceeds 7/8. SiliconANGLE was right that the model didn’t produce a complete answer to the Riemann hypothesis. The README adds a detail nobody printed: the zeta function work was an exception to the standard procedure, and the companion 11/12 proof “was human edited for readability”.

For the rest, OpenAI does state a cost. The model was posed roughly 4,000 problems, and the average result used about three hours of ChatGPT Pro thinking compute. Even so, the company hasn’t named or released the model, and the README warns that “Some of the unformalized results could have issues.” We’ve covered that pipeline before, including GPT-6.1 Astra, which OpenAI pulled from DevDay and delayed indefinitely.

What to watch is that version log. OpenAI says it keeps the public release history, records corrections as new versions, and leaves earlier editions accessible. Two days in, the log already carries three withdrawals, 14 proof repairs and 13 citation updates. Gil Kalai, who matched twenty of the results to problems from his own blog, calls the collection amazing and then adds a third caveat: “even Lean verification may have issues”.