25 Fields Medalists Just Warned the World About AI. The Reason Is Not What You Think.

When machines begin to solve problems: the warning from 25 Fields Medalists and the crisis of mathematical misalignment

On September 11, 2026, a 700-word statement appeared at mathandai.org. The initial signatories included Terence Tao, Peter Scholze, Manjul Bhargava, Maryna Viazovska, Ngô Bảo Châu, and Yu Deng. In total, 25 Fields Medalists signed it. Within days, more than 5,000 people added their names.

A joint statement from 25 Fields Medalists is rare. The last comparable collective action in mathematics came during the crisis over the foundations of the field. Terence Tao's participation matters. He has used AI tools in his own work. His name makes it harder to read the statement as simply anti-AI.

The spark: 88 hours and a Millennium Prize problem

The trigger came three days earlier. On September 8, OpenAI announced that its unreleased next-generation model organized about 10,000 AI agents. In 88 hours, it produced a proof of the existence and smoothness problem for the Navier-Stokes equations. OpenAI released a 166-page paper and Lean formal verification code. The problem is one of the seven Millennium Prize Problems, with a $1 million prize.

What followed was an academic ethics dispute. One day before OpenAI's announcement, New York University mathematician Tristan Buckmaster and Levent Alpoge, who works at Anthropic, published breakthrough results on the same problem for related equations. They had worked for months. Their records were in OpenAI's Codex tool. Buckmaster accused OpenAI of jumping the gun. He said the timeline and method of OpenAI's proof strongly suggested that its model had access to their unpublished research.

Buckmaster also disclosed another detail. The OpenAI project lead offered to make him first author, give him ample computing resources, and share the $1 million prize. The condition: he could not list his collaborator at Anthropic on the paper. Buckmaster refused. He was then asked, "Why are you trying to destroy your career?"

The episode touched the mathematical community's main concerns: priority, credit, and the use of unpublished work.

What the statement says: three lines of misalignment

The statement does not deny AI's technical capabilities. It opens by acknowledging that "in recent months, the mathematical abilities of large language models have improved dramatically, and they have already been able to solve important open problems in several areas of mathematics." The core argument is that the goals of AI companies and the goals of the mathematical community are badly misaligned. That argument has three lines.

The first line: the metric has become the goal. AI companies treat "how many math problems were solved" as a benchmark for model capability. The mathematical community has never pursued getting answers as quickly as possible. The statement says: "Solving problems is only a tool and proxy metric for the primary goal of conceptual understanding and insight." When machines optimize that metric at high speed, "producing 'true/false' statements in bulk at ever-increasing speed may not be nurturing new ideas; it may be destroying the soil in which ideas grow."

The second line: speed destroys understanding. Mathematical knowledge does not move in a straight line from problem to answer. It moves through lectures, discussions, and simplification until graduate students, even undergraduates, can understand it. Some mathematical ideas take decades or centuries before they become widely used tools. Terence Tao used a metaphor: AI companies are "dumping raw meat carcasses on the public dinner table." They care about "creation" and "verification." They abandon "explanation, acceptance, and digestion," three equally critical steps.

The third line: the chain of transmission is breaking. The statement says that if no mathematicians take responsibility for developing AI-generated ideas and integrating them into mathematical knowledge, those ideas "will never truly come alive." The most valuable resources in mathematics are students and ideas. Both require time and human interaction. When AI can produce final results directly, a question follows: if answers can be obtained directly, will humans still go through the training needed to form understanding, ask new questions, and generate new ideas?

Background: this is not the first warning

On June 2, 2026, a "Leiden Declaration" initiated by 16 prominent mathematicians warned about AI's impact on mathematics. More than 600 mathematicians have supported it, and the International Mathematical Union has endorsed it. The Leiden Declaration states that AI is placing the core values of mathematics under threat, with severe effects on students and early-career mathematicians.

The two statements form a progression. The Leiden Declaration was a statement of principles. The statement by the 25 Fields Medalists was an emergency action prompted by the OpenAI incident. Terence Tao wrote on his blog that the signatories considered the problem urgent. They did not wait for a longer consultation process. They published as soon as possible.

That urgency is a signal. The threat felt by the mathematical community has shifted from "possibly in the future" to "happening now."

In the same period, about 1,500 mathematicians began signing a petition. They are boycotting a mathematical marathon sponsored by leading AI companies at Caltech at the end of October. OpenAI withdrew its sponsorship of the event.

A deeper question: who has the right to define "important mathematics"?

The discussion also raised structural problems long present within mathematics itself. On platforms such as Zhihu, a highly upvoted answer asked who has the right to decide what counts as important mathematics. The evaluation system represented by the Fields Medal places the power to define "important work" in the hands of a few. That evaluation affects positions, funding, students, and the careers of young mathematicians.

When AI can rapidly solve problems that would have taken a young mathematician ten or several decades to crack, what should be protected? If what needs protection is human understanding of mathematics, then AI solving problems is not itself a threat. The harder question is whether efforts to limit AI in the name of protecting young mathematicians or the research ecosystem are protecting mathematics or the status and career system of existing mathematicians.

Terence Tao has acknowledged this. He admits that priority competition has always existed among mathematicians. "This was not a problem before, because mathematics was hard enough that a natural braking mechanism formed to prevent things from getting out of control. But now, with no speed limit, everything suddenly begins to collapse."

From a "mathematical crisis" to a "crisis of intellectual labor"

The statement ends with a broader claim: "We are witnessing a general threat to intellectual labor."

That claim extends the discussion beyond mathematics. In many fields, years of professional training serve two purposes: producing final answers or products, and developing the capacity to understand, ask new questions, and generate new ideas. AI systems are built on the body of knowledge humans have accumulated. They are increasingly able to produce the results of such work directly. The resulting misalignment affects mathematics and every domain of intellectual labor whose core value is understanding.

The statement leaves one main legacy: it is a collective protest by mathematicians against AI companies, and also a deeper inquiry into how to protect understanding, the core value of human cognition, in the age of AI.

The statement does not deny AI's potential. It acknowledges that AI "offers the potential to strengthen and accelerate genuine mathematical research and understanding," but stresses that whether these changes are beneficial or destructive "depends largely on the decisions made by those who control the technology."

Mathematics does not need to be "protected" from AI. What mathematics needs is to reaffirm, in an age of greater speed and cheaper answers, its most fundamental pursuit. Its pursuit is understanding the world. Possessing answers is secondary.

No technology has replaced that pursuit.