Terence Tao is not an AI sceptic. For two years he has been the mathematician most willing to sit down with ChatGPT, feed it problems, pair it with the Lean proof assistant and show colleagues what it can do. When Tao says something has gone wrong, the maths world listens.
On September 11, 2026, Tao published a post on his blog with a plain title: “A severe misalignment of AI in mathematics.” Attached to it was a declaration signed by 25 winners of the Fields Medal, the prize often described as the Nobel of mathematics, and a note that the group had written it in a week because they felt the situation could not wait.
The declaration does not say AI is bad at maths. It says the opposite, and that is the problem.
The week that pushed 25 medalists to act
To understand the declaration, you have to understand the four days before it.
Monday, September 8. OpenAI announced that an internal model, running roughly 10,000 AI agents in parallel, had produced a result on the Navier-Stokes equations in 88 hours. Navier-Stokes is one of the seven Millennium Prize Problems, each carrying a $1 million bounty from the Clay Mathematics Institute. OpenAI’s claim was that a smooth fluid, starting at rest and pushed by a smooth force, can reach infinite velocity in finite time, what mathematicians call blow-up. The company said it first ran 1,000 agents on simpler versions of the problem for 50 hours, then pointed the full swarm at the real thing.
The same week. Tristan Buckmaster, a mathematician who has spent years on exactly this question, published a sharply worded statement. After speaking with OpenAI, he said, he found the company’s route to the result was almost identical to the one he and a collaborator had been pursuing. The word plagiarism began appearing in coverage. TechCrunch summed up the mood with a headline calling it an escalating feud between OpenAI and mathematicians.
Also that week. OpenAI’s newly launched GPT-6 Astra was being marketed partly on maths results, including new bounds on gaps between prime numbers and a 97.6 percent score on the FrontierMath Tier 4 benchmark.
Thursday, September 11. The declaration went live at mathandai.org.
Put together, the sequence explains the tone. This was not a philosophical essay written at leisure. It was a response to a company treating the hardest open problems in mathematics as a scoreboard.
What the declaration actually argues
The document is short, about 700 words, and it makes five connected points. Here they are in plain language.
1. AI can now solve serious problems. The signatories accept that large language models have improved dramatically in recent months and can crack major open problems across many fields. They are not disputing the capability.
2. Using famous problems as a benchmark is harmful. The core sentence of the declaration says the push by AI companies to solve mathematical problems as a benchmark damages both the science and the community, and that the two sides’ goals are “severely misaligned.” That phrase, borrowed from AI safety, is the headline.
3. A solved problem is not the point. In mathematics, the signatories explain, a famous problem is a lighthouse. Solving it matters because of the new ideas it forces into existence, which then get discussed, simplified and eventually taught to students. If a machine hands over a proof and nobody does that follow up work, the ideas never enter the canon and the chain of understanding between generations breaks.
4. Rushed announcements raise attribution and plagiarism questions. Results announced without a proper write up, without isolating the new methods, and without citing the people who did the earlier work create exactly the kind of credit dispute now playing out over Navier-Stokes.
5. This is not only about mathematics. The signatories say the same misalignment threatens every profession where years of training were meant to build understanding, not just outputs. They frame maths as an early warning for science, the creative professions and society as a whole.
The declaration closes by saying AI could still enhance real mathematical understanding, that the profession will have to adapt, and that whether the outcome is good or destructive will depend on the humans controlling the technology.
Who signed
All 25 initial signatories are Fields Medalists, spanning nearly fifty years of the prize:
- Pierre Deligne (1978), the most senior
- Simon Donaldson (1986), Shigefumi Mori (1990), Pierre-Louis Lions and Efim Zelmanov (1994), Maxim Kontsevich and Curtis McMullen (1998)
- Andrei Okounkov, Terence Tao and Wendelin Werner (2006)
- Elon Lindenstrauss, Ngô Bảo Châu, Stanislav Smirnov and Cédric Villani (2010)
- Artur Avila, Manjul Bhargava and Martin Hairer (2014)
- Caucher Birkar, Alessio Figalli and Peter Scholze (2018)
- Hugo Duminil-Copin, June Huh, James Maynard and Maryna Viazovska (2022)
- Yu Deng (2026), who received the medal in July
The site invites other mathematicians to add their names, following the model of June’s Leiden Declaration.
The mathematician who went the other way
The declaration is not the whole of the maths community, and one absence is telling.
Jacob Tsimerman, a Canadian mathematician, also won the Fields Medal in July 2026. At the award ceremony he announced he was leaving the University of Toronto to join OpenAI, saying he expected AI to do mathematicians’ work faster and better within a short time. He did not sign the declaration.
Tsimerman’s move and Tao’s declaration are two answers to the same question. One says: if the machines are about to overtake us, go and help build them. The other says: if the machines are about to overtake us, make sure we do not lose what the work was for. Both come from people at the very top of the field, which is why neither can be dismissed.
This is the second warning in three months
The September declaration follows the Leiden Declaration on Artificial Intelligence and Mathematics, published on June 2, 2026. That document was drafted by 16 scholars from 15 universities, formally endorsed by the International Mathematical Union, and has collected more than 2,600 signatures. It focused on transparency, integrity and attribution when AI is used in research, and Peter Scholze publicly backed it, saying the real goal of mathematics is human understanding.
Tao acknowledged in his post that the new declaration skipped the Leiden process’s careful consultation because the signatories judged the moment urgent. Read together, the two documents show a community that spent the spring writing rules and the summer watching a company ignore them.
What the other side says
A fair account has to give OpenAI and its defenders their argument.
The results are real. If the Navier-Stokes blow-up result holds up under review, it is a landmark, whoever or whatever produced it. Mathematics has always celebrated the first correct proof.
Benchmarks drive progress. AI labs argue that hard, verifiable problems are the best possible test of reasoning precisely because a proof is either right or wrong. Without such targets, capability claims become marketing.
Mathematicians are not being replaced, they are being offered a tool. Some researchers, including several who spoke to the newsletter Understanding AI over the summer, see the change as similar to the arrival of calculators and computer algebra, disruptive but ultimately expanding what humans can do.
Adaptation is already happening. Tao himself has run AI assisted proof projects, and the declaration admits the profession must change.
The signatories’ reply is that none of this addresses the specific harm they name: results announced for competitive advantage, before they can be written up, credited or understood.
Why a maths dispute is a story for everyone
It would be easy to file this under academic quarrels. That would miss what the signatories are saying.
Mathematics is the field where AI progress is easiest to measure and hardest to fake. A proof is checkable. That makes it the first place where the collision between commercial AI timelines and the slow process of human understanding shows up clearly. The declaration’s warning is that what is happening to mathematicians now will happen to scientists, doctors, lawyers, engineers and writers next, as models become able to produce the finished output of years of training without the training.
It also lands in a wider moment of unease. In the same week, former Anthropic researcher Jacob Coxon resigned, warning that the leading labs are racing each other without acting responsibly. Days earlier, Senator Bernie Sanders introduced the Ban Artificial Superintelligence Act. The mathematicians are not calling for a ban or a pause. But their complaint, that the race is distorting the purpose of the work, is the same complaint in a different vocabulary.
What happens next
Three things are worth watching.
Whether the Navier-Stokes result survives review. OpenAI’s claim has not yet been through the kind of scrutiny a Millennium Prize result requires. If it fails, the declaration’s warning about rushed announcements will look prescient. If it holds, the credit dispute with Buckmaster becomes the central question.
Whether OpenAI responds. The company has so far framed its maths work as a demonstration of capability. A public commitment on write ups, citation and coordination with researchers would answer the declaration’s main demand.
How many more names the declaration collects. The Leiden Declaration reached 2,600 signatures. If the medalists’ statement gathers similar support, it becomes harder for labs to describe the concern as a minority view.
Mathematics has survived the calculator, the computer and the proof assistant. The signatories are not claiming it will not survive AI. They are asking who decides what survival looks like, and warning that right now the answer is being written in a lab, not in a seminar room.
Frequently Asked Questions
What is the Fields Medalists’ AI declaration on mathematics?
It is a joint statement titled “A Severe Misalignment of AI in Mathematics,” published on September 11, 2026 at mathandai.org and signed by 25 Fields Medal winners. It warns that AI companies’ push to solve famous maths problems as a benchmark is harming mathematics and that the goals of the AI industry and the maths community are misaligned.
Who signed the declaration?
Twenty-five Fields Medalists, including Terence Tao, Peter Scholze, Pierre Deligne, Maryna Viazovska, Martin Hairer, Cédric Villani, Manjul Bhargava, James Maynard, June Huh and 2026 winner Yu Deng. Other mathematicians can add their names on the site.
What triggered the declaration?
OpenAI’s September 8 announcement that an internal model using around 10,000 agents had produced a blow-up result on the Navier-Stokes equations in 88 hours, followed by mathematician Tristan Buckmaster’s statement that the approach closely matched his own unpublished work.
Is the declaration against AI in mathematics?
No. It says AI can now solve major problems and could enhance mathematical understanding. Its objection is to using famous problems as a competitive benchmark, announcing results without proper write ups or credit, and losing the human process by which ideas are understood and passed on.
How is it different from the Leiden Declaration?
The Leiden Declaration, published June 2, 2026 and endorsed by the International Mathematical Union, set out principles for transparency and integrity when AI is used in maths research. The September declaration is a shorter, urgent response by Fields Medalists to specific recent events, and Tao said it was written in about a week.
Did all Fields Medalists agree?
No. Jacob Tsimerman, a 2026 Fields Medalist, left the University of Toronto to join OpenAI in July and did not sign. He has said he expects AI to do mathematicians’ work faster and better.
What does “severely misaligned” mean here?
The signatories borrow the term from AI safety but use it practically: AI companies want fast, headline making results on competitive timelines, while mathematics needs slow verification, proper attribution and conceptual understanding. The two sets of goals pull in different directions.
Why does this matter outside mathematics?
The declaration argues that mathematics is an early example of a problem every knowledge profession will face: AI producing the finished output of years of training without the understanding that training was meant to create.
