Archive · 8 Oct 2026
OpenAI Releases 722 Math Manuscripts, Prompting Calls for Scrutiny
OpenAI shared 722 math papers made with help from a computer model that it has not released.
The papers are arranged into 372 groups of related results.
OpenAI says the model worked on about 4,000 problems and produced solutions or progress on many difficult questions.
Some proofs are written in Lean, which can help a computer check whether the steps follow correctly.
Other papers still need more checking, and OpenAI says corrections will appear in new versions.
Mathematicians will need time to decide which results are correct and important.
A group called the Association for Human Mathematics says mathematicians should stop working with OpenAI.
Its members worry that AI research may overlook human work and scientific practices.
OpenAI says it consulted an independent advisory group and wants to improve how it shares this kind of work.
OpenAI published 722 manuscripts in 372 groups, produced by an unreleased internal model.
The company says the work includes solutions or progress on hundreds of open problems; one report counts 377 new results.
OpenAI says it posed roughly 4,000 problems to the model, with typical results using about three hours of ChatGPT Pro-level compute.
Many proofs have Lean formalizations, but some do not; OpenAI says it will record corrections as new versions.
AHM urged mathematicians to reject collaboration with OpenAI, while researchers have also raised concerns about verification, attribution and publication practices.
- Who
- OpenAI, the Association for Human Mathematics, and mathematicians assessing the work.
- What
- OpenAI released 722 manuscripts of mathematical results generated by an unreleased internal model; AHM criticized the release and urged mathematicians to reject work with the company.
- When
- OpenAI announced the release on Tuesday, October 6; AHM issued its criticism two days later. The year is not specified.
- Where
- The manuscripts were published in a GitHub repository.
- Why
- OpenAI says it is sharing the model's mathematical results and developing practices for communicating AI-generated research; AHM objects to the release's scale and says it conflicts with scientific norms.
OpenAI's case for sharing
Critics' concerns
Value of the results
OpenAI's case for sharing
OpenAI says the collection contains solutions or significant progress on hundreds of open problems and could give researchers new tools.
Critics' concerns
AHM rejects the claim that the release advances mathematics; critics say the results' significance and correctness require independent expert scrutiny.
Release process
OpenAI's case for sharing
OpenAI published the papers on GitHub with revision and citation procedures, Lean formalizations for many proofs, and summaries of reasoning for selected results.
Critics' concerns
AHM says releasing more than 700 manuscripts at once demonstrates power rather than scholarship. Critics also want greater transparency about the model, prompts, computing costs and validation.
AI's place in mathematics
OpenAI's case for sharing
OpenAI says it is consulting an independent advisory group and is developing best practices for releasing AI-generated mathematical research.
Critics' concerns
AHM urges mathematicians to discontinue work with OpenAI, arguing the release risks sidelining human understanding and established scientific norms.
- Manuscripts
- 722
- Result groups
- 372
- Reported new results
- One report says the collection contains 377 new results; OpenAI describes solutions or significant progress on hundreds of problems.
- Problems posed
- About 4,000, according to OpenAI.
- Typical compute estimate
- About three hours of ChatGPT Pro-level reasoning compute per result, according to OpenAI.
- Proof checking
- Many proofs have Lean formalizations, but not all.
- Advisory group
- OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study.
- Verification
- The claims have not been independently verified, and researchers cannot reproduce the original model's reasoning because it remains unreleased.
Sources
OpenAI’s unreleased AI model solves hundreds of long-standing math problems
OpenAI Publishes 722 AI-Generated Math Manuscripts From An Unreleased Model: Here's Why It's A Big Deal
OpenAI releases 372 groups of math results from unreleased frontier AI model