OpenAI's flagship collection of AI-generated math proofs just got its first documented bug report, along with a fix.
Chapter 6 of OpenAI's Ten Advances in Mathematics and Theoretical Computer Science claims an exponential parallel-repetition theorem covering all finite two-player, one-round entangled games. The proof leans on a "greedy conditioning lemma" meant to pick a small set of coordinates so that, after conditioning on winning all of them, a random remaining coordinate is still won with probability at least 1-delta. A new arXiv note finds the printed proof gets its polarities backwards: its test for continuing the argument is written in terms of average success, when the next step actually needs a coordinate with high conditional failure probability. That implication is simply false, and the authors supply a counterexample where the original procedure has no valid next move.
The lemma's own statement, and the parameters the rest of the chapter relies on, come out unchanged, so this is a local repair rather than a takedown of OpenAI's broader claim. Still, it's a useful data point in the argument over AI-generated proofs: the reasoning was plausible enough to pass casual scrutiny while quietly swapping "success" for "failure" at a load-bearing step, exactly the kind of reversal a reader skimming for plausibility rather than checking each implication would miss.
The authors are explicit that patching this lemma is not the same as independently verifying OpenAI's main parallel-repetition theorem.