OpenAI's unreleased AI model posts 719 math papers, splitting mathematicians

OpenAI has released on GitHub a collection of mathematical results produced by an AI model it has not yet made public. Published in early October 2026, the set comprises hundreds of research papers that the company says resolve long-standing open problems. The release has drawn sharply divided reactions from mathematicians.

The collection held 722 papers when it was first announced. On 7 October, three were withdrawn after an error was found in a proof, bringing the total to 719. No concrete objection to the remaining papers has surfaced publicly so far. OpenAI says each result consumed compute roughly equivalent to three hours of ChatGPT Pro reasoning.

Verification is still underway

To check the work, OpenAI relies on Lean, a system that expresses proofs in a formal language so a computer can verify every logical step. Close to half of the papers have been confirmed this way. Independent scrutiny is still needed, however, not least to make sure that the formally checked statement actually matches the claim made in each paper. Scientific acceptance of the results therefore remains pending.

French mathematician Hugo Duminil-Copin, known for his work in probability theory and statistical physics, said during a video conference that 16 problems he had spent years on appear in the collection. In his view, the core problem is not that a machine solved them first, but that the results arrived at enormous scale without giving the community time to absorb them. An earlier open letter signed by 25 Fields Medal winners had warned that releasing many solutions at once leaves too little room to understand the ideas and methods behind them.

Boycott call and dissenting views

The Human Mathematics Association (HMA) went further, arguing that the work is more a show of technological power than a scientific contribution and urging mathematicians to boycott collaboration with OpenAI. Fields Medalist Timothy Gowers takes a different line, saying the sheer number of results should not automatically be seen as negative. His worry is longer term: open problems help draw young people into mathematics, and it is unclear how that pipeline will hold up.

OpenAI said it will fund workshops, conferences and dedicated research programmes to help the community make sense of the results. It is also sharing reasoning summaries, compute estimates and verification materials. The episode marks a new stage in the debate over AI's role in producing science, and over how human researchers can keep pace with it.