Fields Medal winners including Terence Tao and Deng Yu have lashed out, pointing out that AI is ruining mathematics.

The long-standing conflict between the mathematics community and large AI firms has been fully brought to the surface within a week.

September 12, Zhidx News: Terence Tao recently released a post stating that 25 Fields Medal laureates have jointly signed an open letter, slamming AI companies for taking "solving well-known mathematical problems" as the benchmark for measuring model capabilities, describing this behavior as "harm to the science of mathematics and the mathematical community", with the criticism directly targeting OpenAI.

The trigger for this controversy was a high-profile release from OpenAI at the beginning of this month. On September 8, OpenAI announced that its internal model, supported by the concurrent operation of about 10,000 Agents, solved the Navier-Stokes equations problem in only 88 hours. This is one of the seven "Millennium Prize Problems" selected by the Clay Mathematics Institute in 2000, with a reward of 1 million US dollars for each problem.

Just a few hours before the announcement was released, mathematician Tristan Buckmaster from New York University and mathematician Levent Alpöge who works at Anthropic had just published a breakthrough on the same problem, and publicly accused OpenAI of "racing to release results ahead of others".

The dispute over priority, superimposed with the joint condemnation, escalated this incident from a release controversy to a collective statement from the mathematics community against the problem-solving methods of AI companies. The 25 Fields Medal laureates who jointly signed the open letter include Pierre Deligne who won the award in 1978, Ngô Bảo Châu who won the award in 2010, and Deng Yu who just won the award in 2026.

Although Terence Tao did not mention OpenAI directly in his post, he specially attached a report on this controversy from *The Economist*, whose title is "Top mathematicians are furious about OpenAI's methods", directly pointing to OpenAI.

Open letter address:

https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/

01 .

Solving the century-old problem in 88 hours:

The full story of the controversy

According to OpenAI, they used about 10,000 Agents, first ran 1,000 Agents on simplified problems for 50 hours, then fully focused on the Navier-Stokes equations, with a total time of 88 hours. The conclusion they "proved" is that there exists a type of normal fluid that is completely static at the beginning, and under the action of smooth external force, its velocity will rise to infinity in a finite time, which is what is mathematically called "blow-up".

In other words, the Navier-Stokes equations, a continuum medium model, cannot always describe the real world under extreme circumstances. But compared with the conclusion itself, mathematicians are more concerned about another thing: how this proof was made, and whether it is valid.

OpenAI claimed that the model that completed the problem-solving was only trained from August 28, with significantly stronger capabilities than GPT-6 Astra, and is not open to the public for the time being. The whole round of tackling the problem consumed millions of dollars in computing power.

It should be noted that this proof has not yet undergone peer review. According to the rules of the Clay Mathematics Institute, problem solvers must first publish their results in a peer-reviewed journal, and then go through a two-year inspection period to receive the bonus.

The focus of the controversy lies in the issue of authorship. On September 7, the day before OpenAI's announcement, Buckmaster and Alpöge released their results on "related but different" problems such as the forced Euler equations, plus three draft papers, totaling 245 pages. Buckmaster also attached a strongly worded statement, saying that after communicating with OpenAI, he found that the other party's problem-solving ideas were almost exactly the same as the route the two of them were taking.

In academic tradition, rushing to release results after learning about others' unpublished research directions is one of the most taboo behaviors. Buckmaster said directly that the work records of him and Alpöge for several months were left in OpenAI's Codex, so he questioned whether OpenAI had read user data.

Buckmaster said OpenAI replied that the model did not access user data, but avoided talking about the training data issue. Buckmaster also said that OpenAI proposed a "merger plan", in which he would write to announce that "OpenAI's internal model solved the Navier-Stokes problem", on the condition that OpenAI is acknowledged, but Alpöge, who is also a problem solver and works at Anthropic, cannot be listed as an author.

OpenAI completely denied all of this. Sébastien Bubeck, who led the Millennium Prize Problem tackling, said at a briefing: "Neither the researchers nor the Agents saw any of their work before they released it publicly."

He also made public the SMS records, saying that he provided the other party with an option to "publish first"; Sam Altman also came out to defend, saying that Bubeck and the team acted with "integrity and generosity" throughout the process.

However, OpenAI left a thought-provoking supplement in its official blog: "The possibility is low, but we cannot rule out that de-identified data generated by them using our products has helped improve our model."

In the same week, the friction spread to reality. The California Institute of Technology originally scheduled to hold a math hackathon on October 30, allowing participants to use large language models to solve problems, and OpenAI and Anthropic jointly provided 2 million US dollars in AI computing power credits.

After a group of mathematicians released an open letter slamming that such activities are likely to cause destructive impacts on the mathematical community and accusing AI companies of committing academic misconduct, the organizing committee responded that it would add a verification period and public publication requirements. According to a report by Business Insider, OpenAI has now announced its withdrawal from sponsorship.

02 .

Taking problem-solving as the goal

What will mathematics lose

Back to that open letter, all 25 signatories are Fields Medal laureates, with award years spanning nearly half a century. The most senior one is Pierre Deligne, the 1978 laureate, and the latest one is Deng Yu, a Chinese mathematician who just won the award in July this year.

The Fields Medal is awarded every four years, only to mathematicians under 40 years old, and is known as the "Nobel Prize in Mathematics".

The open letter stated that due to the urgent situation, there is no time for a more thorough consultation process, and the open letter was released simultaneously on Terence Tao's blog. Terence Tao was previously known for being open to AI, and has publicly demonstrated using ChatGPT with Lean to do formal proofs on many occasions, but this time he stood on the opposite side of OpenAI.

Although the open letter did not name any company throughout the text, the outside world almost unanimously interprets it as targeting OpenAI. The title of *The Economist*'s report is "Top mathematicians are furious about OpenAI's practices", and summarizes that they warn that AI may destroy the foundation of mathematics.

The core point of the open letter is: In the past few months, the mathematical capabilities of LLMs have advanced by leaps and bounds, and they have been able to solve major unsolved problems in multiple fields. However, AI companies taking problem-solving as a benchmark to promote development is harmful to the science of mathematics and the mathematical community. The goals of AI companies are severely misaligned with the goals of the mathematical community.

The open letter believes that this is only part of a broader misalignment that also affects other sciences, creative industries, and even the whole society.

Famous problems have always existed as "landmarks" and "beacons". Solving such a problem is often accompanied by the emergence of new insights and new methods, followed by a long and arduous process of digestion, presentation, discussion, and simplification, until it is precipitated into textbook content suitable for graduate students and even undergraduates to learn. Some ideas will not become tools used by all mankind until decades or hundreds of years later.

"But problem-solving is only a tool and proxy to achieve the core goal (conceptual understanding and insight)." The open letter warns that forgetting this point will make the tool backfire on the goal: "Mass-producing 'true/false' propositions at an increasingly fast pace may destroy the fertile soil instead of injecting vitality into new ideas."

The open letter believes that many results are released too quickly, leaving no time for standardized manuscript sorting, refining new methods, and citing previous work, which will cause serious authorship and plagiarism problems in the creative industry.

A more long-term concern is the issue of inheritance. If no mathematician is willing to undertake the cultivation work and integrate the ideas conceived by AI into the mathematical classics, these ideas will never really come to life, and the human inheritance chain passed down from generation to generation in the mathematical community will also break.

It is worth noting that the open letter does not oppose AI. It acknowledges that AI has the potential to enhance and accelerate real mathematical research and understanding, and also acknowledges that the profession of mathematics needs to adapt to changes. The open letter leaves this sentence: "Whether these changes will ultimately benefit the field or cause destructive consequences will largely be determined by the decisions of the humans who control this new technology."

This letter did not appear out of nowhere. In the past few months, conflicts between AI and mathematics have occurred one after another: In May, a counterexample from OpenAI's internal model overturned the 80-year-old Erdős unit distance conjecture in combinatorial geometry. The supporting verification paper was jointly completed by nine mathematicians, one of whom is Jacob Tsimerman who later won the 2026 Fields Medal;

On June 2, the *Leiden Declaration on Artificial Intelligence and Mathematics* was released, drafted by 16 scholars from 15 universities, officially endorsed by the International Mathematical Union (IMU), with more than 2,600 signatories;

In July, Alpöge used Claude to falsify the 87-year-old unsolved Jacobian conjecture;

In August, Terence Tao wrote an article judging that AI may make mathematics encounter "the biggest crisis since Gödel";

In September, OpenAI's 88-hour problem-solving pushed all this to a boiling point.

03 .

Even understanding will be skipped:

Ballpoint pens, calculators and a new concern

Mathematicians' criticism of new technologies has many precedents in history. Socrates complained that writing would make people lazy to memorize; when ballpoint pens appeared, some people worried that pens would die out and writing would lose its sensory impact; the popularization of electronic calculators also scared a group of people.

As a result, people still write, just faster, and mathematicians still do research, just calculate faster.

But there is a new anxiety in this accusation against technology, that is, even understanding itself may be bypassed. A 2011 study by Betsy Sparrow's team at Columbia University found that after the popularization of search engines, people will choose to search instead of recalling when encountering difficult problems, and memory is outsourced to the Internet; recent research by Michael Gerlich from SBS Swiss Business School shows that frequent use of AI tools is negatively correlated with critical thinking ability.

Back to mathematics itself, the open letter also mentions a rarely discussed concern — the disappearance of ancillary benefits. The academic circle has always respected the person who first proves a theorem or solves a conjecture, but the final answer is only part of the academic process. Solving a problem will bring up more problems, and then generate new research fields. If AI outputs the proof directly, this process may not happen.

If AI one day starts to conquer cancer and improve energy supply, few people will care how the breakthrough came about. But abstract mathematics is an exception: it has almost no direct practical application, and its main purpose is human understanding itself. If future abstract proofs are completed by machines and humans cannot understand them, the purpose of this undertaking will become difficult to discern.

Teleporting a person to the summit of Mount Everest is completely different from climbing Mount Everest in person.

Attitudes within the mathematics community on this matter are not yet unified. One of this year's Fields Medal laureates, Canadian mathematician Jacob Tsimerman, announced at the July award ceremony that he would leave the University of Toronto to join OpenAI, on the grounds that he judged that AI would soon do mathematicians' work "faster and better". And he did not sign this joint letter.

The following is the full translation of the original open letter:

*A Severe Misalignment of AI in Mathematics*

I am proud to be one of the 25 initial signatories of the following declaration — all Fields Medal laureates. This declaration stems from our discussions over the past week. We have also published the declaration on this web page, and (similar to the Leiden Declaration) welcome more signatures to join. (Unfortunately, we did not have time to go through a more thorough consultation process like the Leiden Declaration; but we judged that the urgency of the situation requires us to speak