OpenAI Withdraws Three Math Preprints After Sign Error in AI Proofs

OpenAI's Ambitious Math Release Hits a Snag

On October 6, OpenAI published 722 preprints in a public GitHub repository, claiming progress on 372 unsolved mathematical problems across geometry, computer science, and algebra. The company touted this as an effort to "push the frontier of human knowledge." However, within 24 hours, the company withdrew three of those manuscripts due to a critical error.

Dan Roberts, a research lead at OpenAI, announced the updates on X (formerly Twitter), noting six new Lean formalizations, 19 modifications, and three withdrawals. The repository now contains 719 manuscripts, with approximately 42% of top-line results formally verified. This rapid correction underscores the challenges of AI-generated mathematical proofs.

The Sign Error That Caused the Cascade

The root cause was a sign error in a formula within a manuscript on the algebraicity of Weil classes on split abelian varieties. Specifically, an incorrect +1 appeared instead of -1 in a geometric operation. This seemingly minor mistake invalidated a cancellation of terms essential to the proof, breaking the argument's logic.

Because mathematical results often build on each other, the error cascaded. Two other papers—one on Kuga–Satake correspondences for K3 surfaces and another on the rational Hodge conjecture for products of K3 surfaces—relied on the flawed result. All three papers belonged to Repository Outcome Family 032, originally titled "Hodge and Kuga–Satake Results for All Projective K3 Surfaces." The family has since been renamed to "The Rational Hodge Conjecture for CM Abelian Varieties."

Corrective Actions and Community Response

OpenAI revised 14 other manuscripts to fix proof errors, clarify theorem statements and hypotheses, and correct one obsolete citation. The company also updated citations for 13 additional papers. Each withdrawn preprint carries a notice explaining the gap, and the original versions remain accessible via archival links for transparency.

Alex Townsend, an associate professor of mathematics at Cornell University, said the cascade was not surprising. "That an error cascaded into three of the OpenAI manuscripts is not surprising," he told Retraction Watch. "I suspect there will be more errors found." His skepticism reflects a broader concern about the reliability of AI-generated proofs.

An OpenAI spokesperson emphasized that the withdrawals pertain to the proofs, not the mathematical propositions themselves, which may still be correct. "We welcome scrutiny and feedback from the mathematical community," the spokesperson said. "Where errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found."

The Bigger Picture: AI in Mathematics

This incident highlights the delicate nature of mathematical verification, especially when proofs are generated by AI. The Hodge conjecture, one of the Millennium Prize Problems, is notoriously difficult, and these papers were among the most closely watched. The retraction demonstrates that even advanced AI models can make subtle errors that human mathematicians might catch.

OpenAI has positioned this release as a step toward accelerating mathematical discovery. The company also attempted to solve about 3,600 additional problems but only released successful proofs. To address concerns, OpenAI promised to pay mathematicians to study AI-produced results, though specific funding details remain unclear.

However, a significant gap persists: ordinary researchers cannot yet use the model that generated these results. Publishing AI-generated papers is one thing; enabling mathematicians to apply the same tools to their own research questions is another. This limitation has led some mathematicians to urge an end to cooperation with OpenAI, as reported by XenoSpectrum.

Why This Matters for AI-Driven Research

The withdrawal serves as a real-world stress test for AI in scientific research. On one hand, the speed of correction—less than 24 hours—shows that AI can quickly identify and address errors. On the other hand, it raises questions about the initial quality control processes and the potential for more errors lurking in the remaining 719 manuscripts.

For the broader tech and investment communities, this episode underscores the importance of rigorous validation in AI-driven fields. As one analysis noted, "For traders assessing risk-to-reward ratios, this underscores the importance of rigorous validation in value investing within cryptocurrency and AI-driven research." While that may be a stretch, the principle holds: AI outputs require careful scrutiny.

OpenAI plans to continue updating the repository with new formalizations and any errata it notices. The company remains committed to improving clarity and presentation of future papers. For now, the mathematical community will be watching closely to see if more errors surface and how OpenAI handles them.

The incident is a reminder that AI, despite its power, is not infallible. In mathematics, where precision is paramount, every step must be verified. As OpenAI navigates this new frontier, the balance between speed and accuracy will be crucial to earning the trust of the scientific community.