Audit of an OpenAI AI-Generated Proof Finds a Reversed Condition, and Publishes a Repair

An arXiv note posted August 3, 2026 by Mikołaj Sienicki and Krzysztof Sienicki reports that the printed proof of a greedy conditioning lemma in Chapter 6 of OpenAI's Ten Advances in Mathematics and Theoretical Computer Science reverses a condition between complementary events, and provides a corrected proof. The authors state the lemma's statement is correct but the printed procedure is flawed, offering a counterexample and a local repair that leaves the lemma's statement and downstream parameters unchanged. The note does not verify the chapter's main parallel-repetition theorem, name the generating model, or include OpenAI's response.

What happened An eight-page note posted to arXiv on August 3, 2026 reports that the printed proof of a greedy conditioning lemma in Chapter 6 of OpenAI's Ten Advances in Mathematics and Theoretical Computer Science reverses a condition between complementary events. The authors say the lemma's statement is correct, supply a counterexample to the printed procedure, and give a corrected proof they describe as a local repair. A short note posted to arXiv on August 3, 2026 by Mikołaj Sienicki and Krzysztof Sienicki reports an error in the printed proof of a lemma inside Chapter 6 of an OpenAI document titled Ten Advances in Mathematics and Theoretical Computer Science. The arXiv listing 2608.14673 is filed under Artificial Intelligence and cross-listed to Quantum Physics, and its metadata describes an eight-page paper with five references. According to the note, the chapter claims an exponential parallel-repetition theorem covering all finite two-player, one-round entangled games — roughly, a statement that the chance of beating such a game falls off exponentially when many copies are played at once by players who share quantum entanglement. The disputed step is what the note calls a quantitative greedy conditioning lemma, used early in the chapter's argument. Its job is to select a small set of coordinates, D, such that after conditioning on the players winning every coordinate in D, a randomly chosen remaining coordinate is won with average probability at least 1−δ. The authors do not dispute that statement; they say it is correct. Their objection is to the argument printed beneath it. As written, the procedure's continuation test is expressed in terms of average success, while the very next step requires the existence of a specific coordinate carrying a large conditional failure probability. The note says that implication is false. To back the claim, the authors give an explicit counterexample, and say that even simple examples can leave the printed procedure with no valid next move — that is, the loop as written can stall rather than produce the coordinate it needs. They then identify what they take to be the intended continuation condition, stated in terms of failure rather than success, and supply a complete corrected proof. They characterize the repair as local: the lemma's statement is unchanged, and so are the parameters the rest of the chapter draws from it, so downstream steps that cite the lemma should not need rewriting. The note is unusually explicit about what it does not establish. The authors write that their correction should not be read as an independent verification of the main parallel-repetition theorem; they fixed one early lemma, not the chapter. Several things are unknown from the source itself. It does not name which model produced the chapter, describe how the text was generated, or say how much human editing it received before publication. It does not report any machine-checked formalization of either the flawed or the corrected proof, does not include any response from OpenAI, and — as an arXiv preprint — carries no indication of peer review. The claim that the printed proof is broken and the repair valid is, at this stage, the authors' claim, open to the same scrutiny they applied. Read the primary source: arxiv.org ↗ https://arxiv.org/abs/2608.14673 Why it matters The reported failure is not a hallucinated citation or an invented number but a single reversed inequality buried in an otherwise plausible argument — the kind of defect that survives a skim and is only caught by line-by-line reading. It is a concrete data point in an unsettled question: how AI-produced mathematics should be checked before it is relied on. Most public discussion of AI errors in technical writing centers on obvious failures: fabricated citations, invented numbers, arithmetic that falls apart on inspection. The defect described here is a different species. A condition stated over one event was written over its complement — success where failure was needed. The surrounding prose reads correctly, the lemma being proved is true, and the parameters all line up. That combination is what makes it hard: the argument is plausible at every level except the one that decides whether it works. Reviewers who check the statement, the constants, and the overall shape can pass it and still miss the break. That has a practical consequence for anyone using models to draft derivations, algorithm correctness arguments, or protocol analyses. Confirming that a result is true is not the same as confirming that the supplied proof establishes it. In this case both things happened to be separable: the claim survived, the reasoning did not. In other cases the same class of slip could support a conclusion that is simply wrong, with no external check to catch it. The workable review unit for AI-produced mathematics appears to be the individual inferential step, not the theorem, and that is expensive human work. The episode also points at formal verification as a discriminating tool. A polarity reversal between complementary events is exactly the kind of defect a proof assistant catches mechanically, because the type of the object produced would not match the type required by the next step. This chapter, as described, was prose mathematics, checked the way prose mathematics has always been checked — by readers. As AI systems produce more candidate proofs than experts can read, the gap between what can be generated and what can be audited widens, and machine checking is the obvious candidate for closing it. Some caution is warranted about how far to read this. It is one lemma in one chapter of one document, reported by two authors in an eight-page note. It supplies no error rate, no comparison against human-written proofs at similar length, and no evidence about whether this kind of mistake is rare or common in AI-generated mathematics. Human-authored papers also carry repairable errors, and the literature has a long tradition of exactly this genre of corrective note. What is new is the subject: the flawed text came from a system, published by a company with a strong interest in how its mathematical output is judged. What to watch next Whether OpenAI acknowledges or corrects the chapter, whether other chapters draw similar audits, and whether anyone independently verifies the main parallel-repetition theorem the corrected lemma feeds into. Also worth watching is whether machine-checked formalization becomes a norm for AI-generated proofs. The first thing to watch is whether OpenAI responds — with an erratum, a revised chapter, or a substantive disagreement. A visible correction would suggest the document is being maintained like a technical artifact rather than a demonstration; silence would leave readers to reconcile the published text against a third-party note. Whether the correction, if made, matches the one proposed here is a separate question worth tracking, since a different repair might not preserve the parameters the chapter later relies on. Second, whether other chapters in the same document draw comparable scrutiny. A single audit is anecdote; a set of independent audits across the ten claimed advances would begin to characterize how reliable the collection is and where the failure modes cluster. The barrier is that this kind of review requires domain experts willing to spend serious time on someone else's output, with little conventional academic reward for doing it. Third, whether anyone independently verifies the main parallel-repetition theorem the corrected lemma supports. The authors were explicit that they did not. Until that happens, the state of the evidence is that one early step now has a proof its authors believe is sound, and the larger claim remains unchecked in public. A machine-checked formalization of the chapter — or even of the lemma alone — would sharply change how much weight the result can bear. Fourth, whether norms form around disclosing and auditing AI-generated mathematics: labeling which arguments were model-produced, stating what human or automated checking was applied, and treating unverified proofs as drafts rather than results. Journals and preprint servers have moved slowly on similar questions. This note is also a preprint and has not itself been peer reviewed, so its own reception — whether specialists in quantum parallel repetition confirm the counterexample and accept the repair — is part of what to follow.