More than 100 open problems solved, OpenAI brings mathematicians into the discussion

As its internal model advances in mathematics and parts of the scientific community raise concerns, OpenAI is working with an independent group of mathematicians to examine how new results should be assessed, communicated, and integrated into research.

More than 100 claimed open-problem solutions

An internal model trained since late August has now reportedly solved more than 100 long-standing open mathematical problems across several areas of the field. That is the figure reported by OpenAI, which links those results to the same system it credits with solving the Navier-Stokes Millennium Prize problem.

Those claims require an important distinction. A solution produced by a model does not automatically become an established mathematical result. It still needs to be verified, understood, placed within the existing literature, and subjected to the field's usual review processes. OpenAI itself says the pace of progress surprised its own mathematicians and prompted internal discussions about how these results should be communicated.

The issue therefore extends beyond whether a model can produce a proof. It also concerns how quickly results can enter a discipline whose validation and transmission mechanisms remain human. Part of the community is challenging the open-problem race

The announcement comes amid a broader dispute. In A Severe Misalignment of AI in Mathematics, published days earlier, a large group of mathematicians criticized the use of open problems as benchmarks for AI systems. Signatories include Artur Avila, Manjul Bhargava, Martin Hairer, Terence Tao, Cédric Villani, and Maryna Viazovska.

Their argument goes beyond whether individual solutions are correct. The declaration warns that mass-producing solutions to known problems could shift attention away from mathematical understanding and toward outputs, while creating additional problems around attribution, prior work, and the time required to integrate new methods into the field.

OpenAI directly references the declaration in its own announcement and acknowledges that these concerns require deeper engagement with the mathematical community. An independent group rather than an internal committee

The Advisory Group on Mathematics and Artificial Intelligence initially includes nine researchers: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. It is hosted by the Institute for Advanced Study, with Camillo De Lellis and Edward Witten listed as the project's principal investigators.

Its scope extends beyond OpenAI. The Institute for Advanced Study says the group intends to advise AI companies on their interactions with mathematical research and the broader community, with particular attention to the responsible presentation and release of mathematical results.

In its work with OpenAI, the group can assess the significance of emerging results, advise on how they should be disseminated, and address academic and professional standards for AI-assisted mathematical research. It will also consider how these tools can support research and education. Internal progress remains outside its remit

The group is designed with several independence safeguards. According to OpenAI, members are not paid by the company, can publish their advice, comment critically on OpenAI's impact on mathematics, and determine their own membership.

One boundary is explicit: the group will not advise OpenAI on how quickly to pursue its internal mathematical progress. Its role primarily begins when new capabilities and results need to be assessed, communicated, or placed in the hands of the research community.

That leaves one of the central questions behind the current debate unresolved: how quickly should AI systems continue working through open mathematical problems? The group can influence how those results enter mathematical research, but not the pace at which OpenAI attempts to produce them.