Read as article
OpenAI Drops 722 AI Math Manuscripts, Igniting Credit Fight
By @sharedot · · 7 pages
- AI
- Mathematics
- Openai
- Lean Proverfer
OpenAI released 722 AI-generated math manuscripts on GitHub, sparking a dispute over credit and demands for verification beyond Lean.
What OpenAI released
OpenAI announced Tuesday that it published hundreds of mathematical results produced by an advanced, unreleased frontier model. Interesting Engineering reports the average result took about three hours of computing, and that OpenAI checked many proofs using Lean, a proof assistant that mechanically verifies the underlying logic. The release follows last month's announcement that OpenAI's models had resolved the Navier-Stokes problem and more than 100 long-standing open problems.
Why the math world is shaken
The scale and speed are the shock: results on problems at the cutting edge of research, produced in roughly three hours of compute each by a model the public cannot access. The SOFX Report notes the claims include a solution to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis — problems that have resisted human mathematicians for decades. Andrew Sutherland, an MIT mathematician, told Scientific American per The SOFX Report that single-agent claims should be considered unverified until the model is released and results can be reproduced. Tristan Buckmaster, an NYU mathematician who worked on Navier-Stokes, suggested AI systems may take human researchers' work and carry it to completion.
The Lean certificates held up — mostly
Startup Fortune reports that OpenAI's Astra model produced ten advances on open problems with machine-checkable Lean 4 certificates posted to GitHub, letting outsiders run the proofs themselves. An August arXiv audit by Mikołaj and Krzysztof Sienicki covering 18 chapter reviews found no surviving confirmed substantive error in a principal result, though it flagged Chapter 8 for major revision and noted some dependencies were only partly checked. The audit stresses Lean's limits: it verifies logical validity, not whether summaries match the formal theorem, whether definitions capture the intended problem, or whether credit was properly assigned. None of the ten results had gone through traditional peer review as of early August, per Startup Fortune.
The attribution fight
The dispute over credit is where the controversy has landed hardest. Startup Fortune reports that Steven Miller, a Yeshiva University mathematician, accused OpenAI's Astra of reusing his 2016 sphere-packing argument without credit, telling Scientific American it seemed "completely systematic" and pointing to research misconduct. An OpenAI spokesperson said the company would "take responsibility for the correctness of these results" and planned minor updates. Startup Fortune also notes an earlier Navier-Stokes claim triggered a 25-signature Fields Medalist rebuke over credit and review. The lesson, per Startup Fortune: "verified in Lean" and "correctly attributed and peer reviewed" are two separate claims, and only one has been settled.
Standards, access and what comes next
An independent advisory board hosted by the Institute for Advanced Study in Princeton has recommended how AI labs should communicate mathematical results, calling for release of prompts and chains of thought, and declaring that public release is "the beginning, not the completion, of the process of human understanding," per Interesting Engineering. Harvard mathematician Melanie Wood said the goal is standards so AI-lab results can advance the field. The board also opposed using advanced problems to test proprietary models and asked for equitable access for the global math community, a concern The SOFX Report says extends to researchers who cannot use the same systems. OpenAI says it plans better citations, explanations and presentation in future releases.