Archive · 9 Oct 2026
OpenAI’s Mathocalypse Raises Questions About AI Proof Verification
OpenAI used an experimental AI model to produce 719 papers about difficult math problems.
Researchers are interested in the results, but they need to check whether the proofs are correct.
Some proofs were translated into Lean, a tool that checks formal mathematical arguments.
Researchers found places where some written explanations did not match the Lean code.
That does not automatically mean either version is wrong, but it means people should compare them carefully.
Mathematicians say humans need to understand and review the work, not just rely on a computer check.
A researcher also warned that better AI might someday create risks for digital money security.
The papers do not show that such an attack is possible now.
OpenAI released 719 manuscripts claiming advances on longstanding mathematical problems, generated by an experimental model.
Only 42% of the proofs were formalised, and only 10 manuscripts included the model’s chain of thought.
Researchers from the University of Cambridge and King’s College London found at least two discrepancies between a written proof and its Lean code.
Mathematicians have stressed that AI-generated results need independent scrutiny and human understanding before others can reliably assess or build on them.
Ethereum Foundation researcher Justin Drake warned of possible future risks to cryptography, but the release does not show that ECDSA or Bitcoin private keys have been compromised.
- Who
- OpenAI; mathematicians and researchers reviewing its mathematical research release; Ethereum Foundation researcher Justin Drake.
- What
- OpenAI published 719 AI-generated mathematical manuscripts, prompting debate over proof verification and a separate warning about possible future cryptographic risks.
- When
- The release followed AGMAI recommendations issued in late September; the articles do not specify when OpenAI published the collection.
- Where
- The work concerns mathematical research; the review included researchers from the University of Cambridge and King’s College London.
- Why
- The scale of the AI-generated results and discrepancies between written proofs and formal code have raised questions about how reliably the findings can be checked and understood.
Potential value of AI results
Need for verification
Publishing AI-generated mathematics
Potential value of AI results
The collection could contain valuable advances, and rapid publication makes the findings available for researchers to examine.
Need for verification
The findings require independent scrutiny; limited formalisation and missing chain-of-thought material make some results harder to assess.
What formal proof checks establish
Potential value of AI results
Lean can mechanically check that an encoded argument follows accepted logical rules.
Need for verification
A checked Lean proof does not guarantee that the accompanying written explanation accurately describes the encoded argument; researchers found at least two discrepancies.
Possible cryptographic risks
Potential value of AI results
Drake urged digital asset holders and the cryptocurrency industry to prepare for possible future risks and develop stronger cryptography.
Need for verification
The threat remains speculative: the mathematical release does not show that ECDSA has been broken or that Bitcoin private keys can currently be recovered.
- Manuscripts released
- 719
- Manuscripts with chain of thought
- 10
- Proofs formalised
- 42%
- Discrepancies identified
- At least two between a written proof and its corresponding Lean code in a paper examining a problem derived from the Navier-Stokes equations.
- Advisory group
- The Advisory Group on Mathematics and Artificial Intelligence (AGMAI) has nine prominent mathematicians.
- Cryptography warning
- Justin Drake described a possible future threat to ECDSA security; the release does not demonstrate a practical attack.