25 Fields Medalists Warn That AI Is Damaging Mathematics
Twenty-five Fields Medalists, among them Terence Tao and Peter Scholze, have signed an open letter warning that the race to show off AI models solving famous math problems is damaging the discipline itself. The statement promptly rose to the top of China's question-and-answer and search boards, setting off a debate about the benchmark culture of the AI industry.
On September 11, twenty-five winners of the Fields Medal — mathematics' highest honor, and a prize awarded only once every four years — published an open letter titled "A Severe Misalignment of AI in Mathematics." The initial signers include Terence Tao, the UCLA mathematician widely ranked among the greatest living mathematicians; Peter Scholze of Bonn; and Deng Yu. Their target is not AI itself, but the way AI companies now use mathematics: as a public scoreboard on which to demonstrate that their models can solve famous, long-unsolved problems.
The letter's argument is that a famous problem is valuable less for its answer than for what solving it teaches. Important open questions function, in the signers' words, as the "landmarks and lighthouses" of the mathematical world: a solution matters because of the new ideas, methods and concepts discovered on the way, which then get reported, debated, simplified and eventually absorbed into graduate textbooks — a process that can take decades. "Solving problems," the statement says, "is only a tool and a proxy metric toward the primary goal of conceptual understanding and insight." Churning out true-or-false conclusions at ever greater speed, it warns, may be destroying the soil in which those ideas grow rather than cultivating it.
Speed creates a second problem: credit. Some AI-produced solutions, the mathematicians write, have been announced in a rush that leaves no time for a proper paper, for distilling what is genuinely new in the result, or for citing the earlier work it rests on — producing serious attribution and plagiarism disputes. And even a correct, novel AI result still needs human mathematicians to develop and integrate it into the body of knowledge; without that, the result never "comes alive," and the chain by which mathematical knowledge passes from researcher to student is at risk of breaking.
The confrontation did not come out of nowhere. As 36Kr, the Chinese tech-news outlet, put it, the accumulated grievances between the mathematics community and the AI industry were laid fully on the table within a single week: Tao has spent days arguing on math forums that AI is damaging the century-old open tradition in which researchers share unfinished ideas with each other, fearing that a swarm of AI agents would race to solve anything made public. A question asking whether the "math arms race" of AI companies should be stopped topped Zhihu's hot list the day the letter appeared, and a Baidu search snapshot ranked the warning third overall — evidence that a dispute over research norms had broken out of the mathematics community and into general circulation in China.
The Weibo discussion also showed where Chinese readers' minds went first. Some framed the letter as mathematicians discovering their livelihood is threatened; a math-education blogger pushed back on that reading: "Who checks whether what the AI produces is even correct? Who does the checking — the AI itself? If not mathematicians, then whom?" He added that what Tao and the others are condemning is companies rushing to claim credit for machine-produced proofs, not the fact that AI can tackle conjectures. A summary of the statement by Yicai, the financial news outlet, was liked more than 1,000 times within hours.
The letter is careful to concede that AI can genuinely accelerate mathematical research. What the signers object to is the pace and the packaging — results announced as benchmark victories before the slow machinery of papers, attribution and teaching can absorb them. That leaves the industry with an uncomfortable question of its own making: if the measure of an AI model is the problems it has solved, what happens when the people who understand the problems ask for a different metric?