25 Fields Medal winners, including Ukrainian mathematician Maryna Vyazovska, have warned that the race of AI companies to solve the most difficult mathematical problems could begin to harm science itself. They are concerned that AI companies are rushing to be the first to announce a breakthrough and doing so before proper verification and clarification of who owns the key ideas.
25 Fields Medal winners, including Ukrainian mathematician Maryna Vyazovska, have warned that the race of AI companies to solve the most difficult mathematical problems could begin to harm science itself. They are concerned that AI companies are rushing to be the first to announce a breakthrough and doing so before proper verification and clarification of who owns the key ideas.
Mathematicians spoke about this in an open letter titled "A Severe Misalignment of AI in Mathematics." It was signed by 25 winners of the Fields Medal, one of the most prestigious awards in mathematics.
The authors acknowledge that models are already capable of helping with very complex problems, so they are not calling for abandoning AI. However, they are concerned about a new race between AI companies. They are increasingly using such problems as a kind of benchmark - a way to show how strong their new model is. According to the authors of the letter, this creates a race in which the main thing is to get the right answer first, while for science it is also important to understand why it is correct, what new ideas it leads to, and how it is connected to previous work.
The authors of the letter argue that this is already causing problems, as results can be announced in a hurry, before a full description of the work is available. This raises an uncomfortable question: where does AI help end and problems with authorship and plagiarism begin?
The biggest risk, they say, is that scientists will start sharing intermediate results less. Researchers now often openly discuss ideas, show rough calculations, and test approaches with colleagues. But if there is an AI company that can spend millions of dollars on calculations and direct hundreds of agents to the same task, the incentive for such openness becomes less. So AI can speed up the delivery of mathematical results, but potentially make the scientific community itself more closed.
The declaration emerged after a conflict over one of the most famous mathematical problems, the existence and smoothness problem of the Navier-Stokes equations.
On September 8, OpenAI announced that its internal AI system had found a solution to this “millennium problem,” which had remained unsolved for nearly 90 years. The company published the mathematical proof and its formalization in the Lean system. According to OpenAI, nearly 100 AI agents worked on the problem in parallel, and the entire experiment, with several large mathematical problems, used about 300 billion source tokens.
But around the same time, New York University professor Tristan Buckmaster and mathematician Levent Alpöge, who works at Anthropic, were preparing their own research in this area. Buckmaster publicly accused OpenAI of pressuring him to include Alpöge among the authors of the paper. He also questioned whether his use of Codex in the research might have helped OpenAI in some way.
OpenAI denied this. The company said the team and AI agents did not have access to the researchers’ unpublished work, and that Buckmaster’s Codex queries could not have influenced the model that produced the result. OpenAI also subsequently withdrew its sponsorship of a math event at Caltech after criticism from the university’s researchers.
Previously, dev.ua wrote that new AI models have already begun to solve complex mathematical problems that until recently were considered too difficult for large language models.
