OpenAI published 722 AI-generated mathematical preprints on October 6, presenting 372 groups of results across algebra, geometry, number theory and mathematical logic. The release prompted debate about whether AI can speed scientific discovery and whether researchers can verify its output at the same pace. Mathematicians also questioned whether the papers explain their reasoning clearly enough for others to understand and build on the results.

The release followed a separate claim in September: OpenAI said an unreleased model had proposed a solution to the Navier-Stokes problem, an open challenge dating back nearly 90 years. The problem asks whether equations describing fluid motion always produce well-behaved solutions. It is one of seven Millennium Prize Problems, each carrying a $1 million prize. OpenAI said the model took about 88 hours to develop its proposed solution, followed by 17 hours of formalisation and verification.

For the October release, OpenAI said it used an internal model to tackle thousands of mathematical questions. It also published formal versions of many proofs in Lean, a programming language and proof assistant that checks mathematical arguments for logical correctness. Formal checking does not replace independent mathematicians’ review of a result’s assumptions, correctness and significance.

The papers remain under independent scrutiny. Retraction Watch reported that OpenAI withdrew three preprints on October 7 after finding a sign error that invalidated an argument in one paper and affected two related papers. OpenAI also revised 14 other manuscripts to address issues with proofs, statements, hypotheses and citations. The corrections intensified questions about how to assess a large volume of AI-generated work before it attracts wider attention.

Mathematicians’ concerns also extend to what research contributes beyond answers: working through hard problems can produce techniques, reveal connections across fields and inspire new questions. AI could assist researchers by testing conjectures, exploring possible proofs and tackling time-consuming problems, while human expertise remains important in evaluating discoveries. OpenAI has announced support for workshops, conferences and other programmes to help researchers examine its findings.

OpenAI published 722 AI-generated mathematical preprints on October 6, presenting 372 groups of results across algebra, geometry, number theory and mathematical logic. The release prompted debate about whether AI can speed scientific discovery and whether researchers can verify its output at the same pace. Mathematicians also questioned whether the papers explain their reasoning clearly enough for others to understand and build on the results.

The release followed a separate claim in September: OpenAI said an unreleased model had proposed a solution to the Navier-Stokes problem, an open challenge dating back nearly 90 years. The problem asks whether equations describing fluid motion always produce well-behaved solutions. It is one of seven Millennium Prize Problems, each carrying a $1 million prize. OpenAI said the model took about 88 hours to develop its proposed solution, followed by 17 hours of formalisation and verification.

For the October release, OpenAI said it used an internal model to tackle thousands of mathematical questions. It also published formal versions of many proofs in Lean, a programming language and proof assistant that checks mathematical arguments for logical correctness. Formal checking does not replace independent mathematicians’ review of a result’s assumptions, correctness and significance.

The papers remain under independent scrutiny. Retraction Watch reported that OpenAI withdrew three preprints on October 7 after finding a sign error that invalidated an argument in one paper and affected two related papers. OpenAI also revised 14 other manuscripts to address issues with proofs, statements, hypotheses and citations. The corrections intensified questions about how to assess a large volume of AI-generated work before it attracts wider attention.

Mathematicians’ concerns also extend to what research contributes beyond answers: working through hard problems can produce techniques, reveal connections across fields and inspire new questions. AI could assist researchers by testing conjectures, exploring possible proofs and tackling time-consuming problems, while human expertise remains important in evaluating discoveries. OpenAI has announced support for workshops, conferences and other programmes to help researchers examine its findings.