Two days after OpenAI's 722 math manuscripts: a sign error forces 3 withdrawals and 14 revisions, and a mathematicians' group urges colleagues to stop working with OpenAI

Why it matters

【Key points】

On Oct 6 OpenAI released 722 math manuscripts from an internal model in the repo openai/math (https://agipulse.org/e/702). Its history shows that on Oct 7 a sign error in "Algebraicity of Weil classes on split abelian eightfolds" forced the withdrawal of that paper and two dependents, including one on the rational Hodge conjecture for products of K3 surfaces, leaving 719. Fourteen more were revised with proof repairs and corrected statements. Lean-formalized top-line results: 300/719, about 42%.

The Association for Human Mathematics said mathematicians did not ask for this work, accused OpenAI of ignoring its advisory group's view that frontier companies should not test advanced problems on internal models, called the 700-file drop "a demonstration of power," and urged mathematicians to stop working with OpenAI. Terence Tao reposted the statement on his blog, which next day ran a guest post on students abandoning math PhD plans. X: about 1,270 posts in 24 hours, a steady 20 to 85 per hour.

【Verifiable】

AHM statement: https://terrytao.wordpress.com/2026/10/07/ahm-statement-on-openais-october-6-release-of-mathematical-documents/

Withdrawals and fixes: https://github.com/openai/math/blob/main/history.md

Guest post: https://terrytao.wordpress.com/2026/10/08/what-should-we-tell-our-students/

【Our read】

1. The withdrawals make the point. Solving costs about three hours of compute per result; checking a proof still takes experts days or weeks. A 700-paper drop shifts that cost onto the field, and 58% of top-line results have no formal proof behind them.

2. The objection is about who sets the pace and who pays for verification, not whether AI can do math. Expect the fight to move to publication rules.

3. Watch for per-paper verification labels, rising Lean coverage, and journal or preprint rules for bulk AI manuscripts.