OpenAI Posts 372 AI-Generated Math Proofs on GitHub

OpenAI published 372 new mathematical results on GitHub, a platform for storing code and project files. Almost all of them were generated by a single AI agent from a single prompt. Each proof cost about three hours of ChatGPT Pro compute on average. The set includes progress on the Riemann hypothesis, a famous unsolved problem in number theory, along with improvements to computational algorithms. Some of the proofs come with formal verification in Lean, a programming language that checks mathematical logic step by step. In response, 25 Fields Medalists, recipients of the highest award in mathematics, signed an open letter opposing the practice. They warn that a flood of finished results with no explanation burns up the ground needed for new ideas instead of helping it grow.

Three hours of model compute for one discovery is cheap. Some of these results do rest on formal verification in Lean rather than just the agent's say-so. But most of the 372 results came out with no explanation at all. That repeats the same problem we see with unreviewed AI-written code. The result might be correct, but trusting it without anyone checking is still risky. Formal verification is the real check here. Everything else is, for now, a raw, unverified result.