Twenty-five Fields Medalists, including Princeton's June Huh and UCLA's Terence Tao, issued a joint warning on March 11 that the AI industry's race to solve famous unsolved math problems risks damaging the very fabric of mathematical research. The statement, published on Tao's personal website and the "Mathematics and AI" site, argues that treating problem-solving as a benchmark for AI performance undermines conceptual understanding, insight, and the education of future mathematicians.
The declaration follows a series of rapid advances by AI companies. On March 8, OpenAI announced its models had solved the Navier-Stokes equations problem. In July, the company said its latest model had proved the Cycle Double Cover Conjecture. Mathematicians now assess that large language models have reached a level where they can tackle important open problems across multiple areas of mathematics.
Tao said the declaration was drafted over the past week through discussions among mathematicians. "Given the urgency of the situation, it was released before a fully exhaustive consultation process could be completed," he explained.
The problem is not the solving - it's what gets lost
The statement draws a sharp line between producing correct answers and doing mathematics. The essence of mathematical research, the signatories argue, is not simply obtaining solutions. Major unsolved questions drive the emergence of new concepts and methods. The long process of discussing, systematizing, and transmitting those ideas to later generations is the discipline's core.
"In many ways, the mathematical community is a small microcosm of humanity," Tao said. "Our most precious resources are students and ideas." He explained that problems are given to students not merely to get answers, but to build research skills through the act of solving. New ideas develop through lectures, discussions, and careful writing - processes that require time and human interaction.
The mathematicians also flagged growing problems as AI-generated solutions appear at speed: inadequate paper preparation, poor organization of new methods, and insufficient citation of prior research. They warned this will lead to disputes over credit and plagiarism, mirroring tensions already seen in other creative fields.
A call for human integration, not rejection
The Fields Medalists do not oppose AI use. They acknowledged that, applied appropriately, AI could strengthen and accelerate mathematical research and understanding. The critical requirement is that new ideas produced by AI must go through a process where human mathematicians examine them and integrate them into the existing mathematical framework. Without this, the transmission of knowledge to the next generation weakens.
Um Sang-il, CI of the Discrete Mathematics Group at the Institute for Basic Science, called the statement "very timely." He said, "I hope that using LLMs to better understand mathematical structures and methodologies will help advance mathematical research, but at present, some AI companies and individuals seem to care only about claiming that they have solved unsolved problems through LLMs, while neglecting the process of achieving deep understanding and further development."
Kwon Hyun-woo, a PhD candidate in applied mathematics at Brown University, said, "There are many people who do not agree with the statement, but I think it is the declaration that has garnered the broadest support among those recently issued in the mathematical community."
The signatories stressed that the current problem extends beyond mathematics. It is a challenge facing other scientific and creative fields, as well as society as a whole. Whether AI benefits mathematics or causes destructive harm, they argued, depends on choices made by the humans who develop and govern these technologies. They urged the mathematical community, AI companies, and society to urgently discuss the issues involved.
Why this matters for science and research professionals
The mathematicians' warning highlights a tension that now runs through every research discipline. When AI can produce answers faster than humans can verify them, the incentive structure shifts. For researchers in any scientific field, the statement underscores a practical reality: tools that accelerate output can also erode the slow, collaborative processes that produce genuine understanding. Professionals working with AI for Science & Research face the same balancing act - using these systems to augment work without letting the speed of generation outpace the rigor of review. The choice, as the Fields Medalists frame it, is not whether to use AI but how deliberately to govern its role in the research pipeline.
Your membership also unlocks:
