OpenAI posts 722 AI-generated math manuscripts, drawing Fields medalist backlash
OpenAI published 722 AI-generated math manuscripts to GitHub on Oct. 6 — organized into 372 results spanning 17 areas including number theory, algebraic geometry and theoretical computer science, with 10 model reasoning summaries and partial Lean formalizations — according to TMTPost. OpenAI said the model had attempted roughly 4,000 research problems at a cost of about three hours of ChatGPT Pro "thinking" per result, and the next day it withdrew three preprints and revised 14 others, cutting the total to 719. Cambridge Fields medalist James Maynard described a "tsunami" heading for mathematics, while Terence Tao and 24 other Fields medalists issued a statement criticizing the mass release for bypassing peer review; Imperial College's Kevin Buzzard said only 1 of 30 papers in his number-theory area passed Lean verification in his sample.
It exposes the gap between how fast AI can generate results and how fast humans can verify them, and may push journals and labs to rewrite submission and verification rules.