OpenAI Releases 700 AI Proofs, Rattling Math Community

On October 6, 2026, OpenAI published more than 700 preprints on GitHub, alleging it had resolved hundreds of long-standing mathematical problems. The massive dump follows rapid AI advances starting in mid-2026, including claimed breakthroughs on the Unit Distance Conjecture, the Navier-Stokes equations, and ten proofs presented alongside OpenAI's Astra model. Among the new papers, the company addressed a weaker variant of the Riemann Hypothesis, known as quasi-RH, and tackled problems spanning probability theory, algebra, and geometry.

The release provoked an overwhelming reaction from academia, with the blog Proofs and Prompts gathering over 100 responses from scholars. Ben Green noted that the preprints resolved roughly three-quarters of the research goals funded by a European grant he received just four months earlier in June. However, many papers drew heavy criticism for poor legibility, prompting PhD student Tristan Humbert to characterize the writing as unreadable slop. Fields Medalist Peter Scholze urged patience while warning that similar AI capabilities could eventually threaten digital infrastructure by breaking standard cryptographic algorithms.

For professional mathematicians, the sudden arrival of machine-generated proofs fundamentally alters how research is conducted and shared. Terence Tao lamented the lack of human collaborators to explain or teach the new ideas, while 2022 Fields Medalist Hugo Duminil-Copin described feeling paralyzed after OpenAI targeted every major open problem he had publicly discussed. Prior to the release, 25 Fields Medalists had issued a statement highlighting a growing rift between commercial AI goals and scientific practice, and the Association for Human Mathematics called for a boycott of the tech firm.

While some researchers reported using Astra collaboratively or leveraging tools like Anthropic's Claude, others worry about the burden placed on academics. Experts point out that researchers are now forced to spend time deciphering opaque preprints without clear authorship or failure rates. Rather than spending years gradually uncovering mathematical structures, practitioners face an environment where frontier models use unsolved conjectures as test beds, forcing human academics to pivot from discovering proofs to evaluating raw AI output.