Mathematicians urge colleagues to stop working with OpenAI

OpenAI published 372 results from an internal model it has not released, which it says settle or move forward long-standing open problems in maths and theoretical computer science, Scientific American reports. The claims include the four-dimensional Kakeya conjecture, and many proofs have been checked in Lean, a language that verifies a proof’s logic. A spokesperson said one prompt to one AI agent produced almost all of them.

The dispute is about disclosure. An advisory group of mathematicians, formed after OpenAI’s earlier Navier-Stokes result, recommended publishing the model, the exact prompt and the compute time behind each result. OpenAI gives only average compute time and no prompts, and says the recommendations do not bind it. MIT’s Andrew Sutherland told the magazine that one-shot claims stay unverified until outsiders can replicate them; Toronto’s Daniel Litt said the release is good for mathematics.

For buyers weighing a lab’s research claims, anyone can rerun a Lean check on a proof, while how a model produced it stays unverifiable without the model and the prompt.