OpenAI Pulls 3 Math Preprints a Day After 722 Release

One day after posting 722 AI-generated math preprints, OpenAI withdrew three manuscripts over a sign error that cascaded into dependent results.

@sharedot

OpenAI Pulls 3 Math Preprints a Day After 722 Release

By @sharedot · · 7 pages

  • AI Frontier
  • AI Mathematics
  • Openai

One day after posting 722 AI-generated math preprints, OpenAI withdrew three manuscripts over a sign error that cascaded into dependent results.

What happened

OpenAI published 722 preprints on October 6 in a GitHub repository, claiming progress on 372 problem families spanning geometry, computer science and algebra, produced by an unreleased internal model.

According to Retraction Watch, on October 7 the company withdrew three manuscripts because of a sign error — one invalidating an argument and two dependent papers — and revised 14 other manuscripts with proof repairs, corrected statements, clearer hypotheses and one citation fix.

Why it is surprising

The reversal came less than 24 hours after a release the company framed as an effort 'to push the frontier of human knowledge.'

The Rundown AI reports OpenAI claims average compute of roughly three hours of ChatGPT Pro thinking per result, including a quasi-Riemann hypothesis proof whose formalization excludes the paper's later applications.

OpenAI itself warns that unformalized results may contain issues, and a single sign error knocking out three papers shows how fragile unverified output is.

The evidence question

Only some results were Lean-formalized, meaning machine-checked step by step; a company spokesperson told Retraction Watch that roughly 50% of the results were released unconfirmed, on advice from its Advisory Group on Mathematics and Artificial Intelligence.

Cornell mathematician Alex Townsend told Retraction Watch he expects more errors and argued OpenAI should have announced Lean-verified manuscripts first, separating them from work needing community help to verify.

The stakes for mathematics

If the results hold up, The Rundown AI notes, mathematical discovery could scale with computing capacity, shifting human effort toward choosing worthwhile problems and judging significance.

But trust is strained: Retraction Watch reports over 8,000 researchers endorsed criticism of the September 8 Navier–Stokes announcement, and the Association for Human Mathematics called the mass release 'a demonstration of power,' urging mathematicians to discontinue work with OpenAI.

MIT's Andrew Sutherland called the quick withdrawals 'the responsible thing to do' but said much more will be needed to earn back trust.

What comes next

The Seoul Economic Daily reports OpenAI set up an independent advisory committee including prominent mathematicians and says it will follow AGMAI recommendations on release protocol, while refusing the field's request to stop testing advanced problems.

OpenAI says it is working toward releasing the model itself, which would let researchers test whether the approach transfers to their own questions. Company research lead Dan Roberts said on X the repository will continue to be updated with new formalizations and errata.

Sources

  1. therundown.ai › OpenAI releases 722 math papers from an internal AI model
  2. retractionwatch.com › OpenAI withdraws three preprints a day after releasing 722 manuscripts on unsolved math problems
  3. en.sedaily.com › OpenAI Posts 722 Math Papers, Cracking 300-Plus Open Problems

More on AI Frontier