AINEWS
Roughly 25 Fields Medal winners, including Tao, are asking AI labs to slow down. The proofs check out. The community that has to read, check, and teach them cannot keep up.
OpenAI published a writeup this summer claiming an AI had produced a machine-checkable solution to the Navier–Stokes problem, one of seven Clay Millennium Prize Problems, each carrying a $1 million purse. Within days, roughly 25 Fields Medal winners, including Peter Scholze, Pierre Deligne, Deng Yu, Cédric Villani, and Terence Tao, signed a public letter asking AI labs to slow down. The proofs check out, the letter says. The community that has to read, check, and teach them cannot keep up.
Tao, widely considered the leading figure in modern mathematics, gave the bottleneck a name in a pair of July 2026 public appearances. He called it "Proof Indigestion" in a SAIR speech covered by QbitAI, and laid out the diagnosis in an ICM 2026 lecture titled "Mathematics in the age of AI". The Fields Medal, awarded every four years to up to four mathematicians under 40, is the field's highest honor; that 25 of them signed the same letter is the unusual fact.
The OpenAI writeup said the system that produced the solution was "an internal model that is significantly more capable than GPT-6 Astra." The Lean proof the model produced, covered by Quanta Magazine, is a binary artifact: a computer can verify it cheaply, but a human has to write the seminar, the textbook chapter, and the lecture that puts a new lemma in a student's hands. Lean is a programming language for formal mathematics, and the verifier's job is to be strict; the reader's job is to understand. When the production rate of verified proofs outruns the rate at which the community can explain them, the canon fills with true statements nobody owns.
The letter names three failure modes. The first is Proof Indigestion itself. The second is data contamination: training a next-generation math model on the output of today's models means training on proofs humans, by design, do not need to read. Tao borrowed a biological analogy for the SAIR speech, reported by QbitAI: breeding mice whose immune systems you never expose to a real pathogen produces a population that cannot survive contact with the wild. Iterative training on human-unreadable proofs, he suggested, produces models that look strong on benchmarks but cannot reach beyond them.
The third is the loss of what the letter calls "cognitive friction." Real mathematical research is the slow business of formulating a question, attacking a near miss, and then letting the failed attempt suggest a better question. The joint letter warns that turning mathematics into a True/False benchmark rewards skipping that friction. The reward function optimizes away the part of the work that generates the next generation of problems.
In the same QbitAI summary, Tao's earlier positions are quoted: a March 2026 interview in which he called the moment a "Copernican revolution" in mathematics, and a 2026 IPAM workshop remark that AI in math and theoretical physics was "ready for primetime." The July letter is not a retraction. It is a request to design the publishing and review pipeline so the capability does not outrun the institution that turns proofs into knowledge.
The letter does not name a throttle speed, but it points at the moving parts: the rate at which preprint servers accept machine-formalized proofs, the rate at which journals commission human-written explanations to accompany them, and the rate at which graduate programs train students to read and extend Lean-style arguments. Each of those is a lever. The coalition is asking labs to coordinate with the people who run those levers before the next big result lands.
The next test is whether the next lab result, whichever it is, arrives with a human-readable companion in the same paper, not six months later in a separate preprint. If it does not, the math community will have to build the throttle itself.