OpenAI Forms Independent Math Advisory Group AGMAI to Review Model Results

OpenAI announced a collaboration with an independent mathematics advisory group called the "Advisory Group on Mathematics and AI" (AGMAI), hosted by the Institute for Advanced Study. One of its current tasks is to recommend ways to evaluate, review, and publish the large volume of mathematical results produced by OpenAI's internal models.

This arrangement addresses a problem more difficult than "can a model prove a problem": when AI generates mathematical results faster than peer review and academic dissemination can handle, how should the mathematical community absorb these results?

The background is that OpenAI claims it began training a new internal model on August 28, and that this model has solved more than 100 long-standing open problems spanning multiple areas of mathematics, with progress so rapid it surprised even mathematicians within OpenAI. But an important qualifier must be retained here: "solved more than 100 open problems" is currently OpenAI's statement about internal research results; the relevant results have not all been made public, nor have they all completed independent review by the mathematical community. The advisory group's own wording is equally cautious—these results are "reported" by OpenAI to have been produced by its internal models.

On September 8, OpenAI published a paper and Lean formal proof for the Navier-Stokes problem. According to the company's explanation, its internal system constructed a solution that forms a singularity in finite time, thereby proving cases C and D in the official Clay Mathematics Institute formulation. OpenAI also stated it has no intention of applying for this Millennium Prize.

This is one of the few results that has already been made public and can therefore be directly scrutinized by outsiders. But public does not equal accepted: from generation to becoming a recognized mathematical result, a proof must still undergo expert verification, supplementary argumentation, tracing of related work, and public discussion. The solution to the Navier-Stokes problem is still under scrutiny by the mathematical community, and whether it will ultimately be recognized as correct remains undetermined.

The role of formal tools here needs to be accurately understood. Even if tools like Lean can check the derivation process, they cannot alone answer whether a result is truly novel, which existing ideas it employs, or how to transform the proof into knowledge that mathematicians can understand and continue to use. The faster a model solves problems, the more easily this seemingly traditional academic process becomes a new bottleneck.

Before this collaboration appeared, the mathematical community had already publicly expressed unease. On September 11, several Fields Medalists participated in publishing an open letter titled "Serious Misalignment of AI in Mathematics," criticizing AI companies for treating open problems as capability benchmarks and arguing that this competition could bring negative effects on mathematical research and the mathematical community.

The core of the controversy is not about keeping AI out of mathematics. The open letter acknowledges that AI could accelerate genuine mathematical research and understanding; what mathematicians worry about is that the continuous pursuit of "how many more problems have been solved" could let the quantity of results overshadow conceptual understanding.

This concern has a realistic basis in academic processes. An important result usually needs to mature gradually through reports, discussions, and repeated revisions: researchers identify new methods within it, place it into the existing body of knowledge, clarify its relationship to prior work, and then organize complex proofs into forms that peers and students can grasp. If solutions appear in concentrated batches, the academic community will find it difficult to complete this work in time; hasty publication can also easily trigger disputes over attribution, citation, and academic credit.

Open problems also have educational value in mathematical research. Many problems are carefully selected and used to train young researchers' judgment and creativity. When they are converted in bulk into model test sets, the original research ecosystem is also affected. OpenAI directly responded to this open letter in its announcement and acknowledged that AI companies need more prudent communication with the mathematical community. AGMAI is stepping forward precisely against this backdrop.

The advisory group's first batch of members totals nine people, including François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. According to the arrangement announced by OpenAI, the advisory group will help judge the importance of new results, discuss how results should be released in a coordinated manner, and provide opinions on academic and professional norms that mathematical research should follow; it will also pay attention to how AI tools can support mathematical research and teaching.

The advisory group's website provides a clearer independence arrangement: members do not accept payment from AI companies, and recommendations directed at AI companies will be published publicly; its service targets are not limited to OpenAI, and as long as a company's model may have a significant impact on mathematics, the advisory group is willing to provide advice. The organization can adjust its members on its own and invite the mathematical community to submit opinions on the release of OpenAI's batch of results.

But this independence also has clear boundaries. The advisory group has no decision-making power over any AI company, and ultimate responsibility still rests with the company; it also will not participate in determining the pace of development of OpenAI's internal mathematical capabilities. In other words, what mathematicians gain is space for speaking, reviewing, and public criticism, but no veto power over the release of results, and even less ability to control the pace at which models continue to advance research. How much effect this arrangement can have will depend on whether OpenAI adopts the recommendations and whether, when disagreements arise, information can be fully disclosed.

In the past, when people discussed AI and mathematics, attention often focused on whether a model could prove a problem. Now, the center of the question is beginning to move backward: who should verify a machine-generated proof, how verification work gains academic recognition, how related ideas are traced to their sources, and in what order concentrated results are released—these will all affect whether the mathematical community can digest this leap in capability.

If handled poorly, researchers may be overwhelmed by a flood of proofs; if handled well, AI has a chance to become a research tool that expands mathematical understanding. The establishment of AGMAI shows that open problems can no longer continue to serve as a simple benchmark of model capability: behind every "solved" hard problem are connections to research traditions, talent development, and knowledge dissemination. Answers appearing faster does not automatically bring deeper understanding.

This advisory group is only the first step. The real test will occur after the first batch of results is gradually made public: whether the mathematical community can independently verify these proofs, whether researchers can extract understandable and reusable new ideas from them, and whether AI companies are willing to leave enough time for rigorous academic processes. If OpenAI's judgment on the more than 100 open problems is ultimately confirmed, mathematical research will face a rare speed shock—at that point, the scarcest resource may no longer be the ability to give answers, but the time for humans to understand, review, and absorb those answers.