OpenAI Withdraws Three of 722 AI-Written Math Papers as Mathematicians' Group Urges Boycott

On October 6, OpenAI published 722 manuscripts on GitHub. An unreleased internal model generated them, and they cover 372 families of related results. OpenAI says the model attempted about 4,000 problems, with most results costing roughly three hours of ChatGPT Pro thinking compute. The headline claims are large in scope. They include a proof of the Unique Games Conjecture, a zero-free region for the Riemann zeta function where the real part exceeds 7/8, progress on the rational Hodge conjecture for CM abelian varieties, and a resolution of the free group factor problem. Each of these is OpenAI's claim, and none has been independently verified.

The next day, OpenAI withdrew three manuscripts, according to Retraction Watch. One paper used +1 where its own conventions required -1. That error invalidated its argument and the construction that two dependent papers relied on. All three concerned the Hodge conjecture. OpenAI also revised 14 other manuscripts with proof repairs, corrected statements and clearer hypotheses, which brought the repository down to 719 papers. According to the repository's own catalogue, Lean formalizations cover the main result of 162 papers. The other results rest on unformalized prose that someone has to referee.

The Association for Human Mathematics (AHM) criticized the release and urged mathematicians to stop working with OpenAI. The group called the release "not a demonstration of scholarship, but a demonstration of power." AHM argues that OpenAI ignored an advisory group's statement that frontier labs should not test advanced open problems on internal models. Not every commentator agrees. A Neowin opinion piece argues that the boycott case is weak, and some AI researchers on X have described the reaction as an attempt by mathematicians to protect their own professional standing.

The withdrawals do not show that the corpus is worthless. They do show where the cost lies. Generating a manuscript took about three compute-hours, but refereeing a claimed proof of a major conjecture takes expert time measured in months. The formalized subset is the only part a reader can trust without that work. OpenAI benefits if the unformalized majority is read as equivalent to the formalized part, and the withdrawals are evidence against that reading.