OpenAI releases hundreds of mathematical manuscripts amid verification concerns

OpenAI released hundreds of AI-generated mathematical manuscripts online on October 6, opening a collection of claimed results to scrutiny while independent advisers stressed that publication does not complete their assessment. The release gives mathematicians access to papers and supporting proof materials, but the company says the results are at different stages of verification.

The Guardian reported concerns about verification and access to the underlying models on October 7. The Advisory Group on Mathematics and Artificial Intelligence, or AGMAI, says the mathematical community must assess the work. Its advisory role, the group emphasized in its October 6 response, neither judges the results’ impact nor endorses the process used to obtain them.

What OpenAI’s 722 manuscripts contain

OpenAI’s public repository lists 722 manuscripts organized into 372 families. A family can include a principal result, companion arguments, consequences or alternative proofs. Those figures describe the collection’s structure: the manuscript total is not a count of separately solved mathematical problems, and the family total does not establish independent acceptance of the results.

The company says the manuscripts and supporting proof artifacts were produced by an internal model. It expanded evaluations on open research problems after performance on its existing mathematical tests saturated. Some outputs build on earlier results produced by the models, meaning parts of the collection depend on other model-generated work.

Which mathematical proofs have been checked?

The repository includes Lean formalizations, which allow computer checking of mathematical proofs, but OpenAI does not describe every manuscript as formalized. Its README explicitly warns: “Some of the unformalized results could have issues.” The company says it will address problems and add further formalizations as they become available.

Readers can use the openai/math repository’s overview to navigate the families and its manuscript map to locate individual papers and supporting materials. The preprints directory contains PDFs, source files and manuscript-specific instructions. A separate formalization catalogue describes available formal proofs, associated papers and verification configurations.

OpenAI also promises versioned corrections, with earlier releases remaining accessible. Its account acknowledges some variation in preparation: while the vast majority of results followed the same procedure, one write-up concerning the Riemann zeta function was edited by a human for readability. These disclosures describe production and checking arrangements rather than establish mathematical acceptance.

Why advisers want human understanding and model access

AGMAI’s October 6 response describes publication as the beginning of understanding and incorporating the results into mathematical knowledge. It argues that “the future of mathematical research cannot consist only of understanding results produced by AI labs.” Mathematicians, it says, need freedom to formulate their own questions and approaches, backed by equitable access to powerful tools and sufficient computing resources.

The group’s earlier September 29 recommendations, informed by more than 600 survey responses, asked laboratories to stop testing advanced mathematical problems on proprietary models inaccessible to the broader scientific community. Those recommendations preceded this release; they set out a wider position on how AI-generated mathematics should enter scholarly work.

The guidance distinguishes papers understood by a responsible mathematician from outputs not yet understood by their human prompters. For the former, it recommends conventional preprint, journal-review and seminar practices. For the latter, it calls for readable proofs, attribution of earlier ideas, repositories outside laboratory control, persistent identifiers and recorded revisions, alongside disclosures about models, prompts and problem selection.

AGMAI says it operates independently, its members accept no payment for this work and it has no decision-making authority at AI companies. Its members include Timothy Gowers, Martin Hairer, Ravi Vakil and Melanie Matchett Wood. The group calls its discussions with OpenAI constructive but leaves assessment of compliance with its recommendations to the mathematical community.

What OpenAI has promised next

OpenAI says it is working toward releasing the internal model, without specifying a date. It also promises funding for workshops, conferences and special programs to help people understand major AI-produced results, with program details to follow. Its announcement commits to improvements in citations, exposition and presentation in future releases.

The advisers’ recommendations warn that restricted models and unequal computing access could deepen disparities between institutions. They call for support for human understanding, including workshops, working groups, students and postdoctoral researchers, with funding decisions handled by existing independent nonprofit institutions. Their access proposals concern the global mathematical community, extending the debate beyond checking the current collection.