Particle.news

OpenAI Publishes 722 AI‑Generated Math Manuscripts, Prompting Worldwide Verification

The release forces mathematicians to check claims because only a minority of results have machine‑checked formal proofs and the underlying model and prompts remain private.

Overview

  • OpenAI posted 722 AI‑generated manuscripts on GitHub on Tuesday that it says come from an unreleased internal model and that together form 372 families of related results.
  • Reports differ on formal checks but agree they are a minority: one catalog lists 162 papers with Lean‑formalized main results while other accounts put the share at roughly 40–42 percent, leaving most papers unformalized.
  • Hours after the drop, OpenAI withdrew three manuscripts on October 7 after finding a sign error that undermined proofs in two related papers and the public catalog is now reported at about 719 items.
  • The release has drawn strong criticism from many leading mathematicians and groups including a Fields‑medal open letter and an Institute for Advanced Study advisory panel, and the Clay Mathematics Institute and journals will adjudicate any prize‑relevant claims through standard peer review.
  • Independent human verification and community formal checks are under way and are expected to take months because key information—full proofs in Lean, prompts, complete chains of reasoning, and training‑data provenance—has not been released.