In a recent SAIR talk, Terence Tao publicly criticized AI companies for relentlessly accelerating without understanding the consequences, calling for a slowdown. He warned of 'proof indigestion' and knowledge congestion as AI rapidly generates and verifies proofs while human comprehension, peer review, and textbook writing lag behind, risking loss of mathematical creativity and data pollution.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
Half a year ago, Terence Tao was still saying AI was "ready for primetime." In his SAIR talk, he publicly asked model companies to slow down. His reason is not that AI cannot produce proofs, but that AI is sprinting ahead only in the first two stages of the proof lifecycle while the last three are nearly stalled—this constitutes a knowledge congestion, and OpenAI's response has precisely sidestepped mathematics itself.
In March 2026, Terence Tao publicly urged colleagues at the IPAM conference not to be too rigid, saying that current AI in mathematics and theoretical physics was already "ready for primetime" and could save a great deal of wasted effort; in an interview with MIT Technology Review the same month, he asserted that mathematics was undergoing a Copernican revolution. In an earlier interview he also said that in the future even middle school students would be able to contribute to frontier mathematical research with the help of AI + Lean.
By the latest SAIR talk, his wording had shifted completely: AI companies accelerate endlessly yet know nothing about what will happen afterward, and "must slow down. There is absolutely no reason to go this fast, not one reason." From March 2026 to this talk, the span of the attitude shift is about half a year.
It should be noted that this article is based on a secondhand relayed report; neither the original text of Tao's talk nor the original text of OpenAI's official statement is directly given in the source, and the specific date of the talk and the full name of the SAIR conference are also not made clear. Therefore the reconstruction of his argumentative structure below should be regarded as an organization of what the report states, not as independent verification of the full text of the talk.
The analytical framework Tao offered is this: the normal complete lifecycle of a mathematical proof usually has five stages—proposing a solution, mechanical verification (such as passing a check in the Lean formal proof language), interpretive work, peer review, and writing into textbooks.
His judgment is that AI models today are accelerating madly only in the first stage, "generating solutions," and the second, "Lean formal verification," while the latter three stages—understanding, digestion, writing textbooks—have progressed almost not at all. The report describes it this way: AI keeps crushing major conjectures, but no one gives a talk, no one writes a paper, and even the person who presses the enter key to generate the proof cannot understand what just happened.
The key to this framework is that it splits "whether AI can produce a proof" and "whether the mathematical community can absorb this proof" into two separate questions. The former has already been solved by AI; the latter has not. Tao called the resulting knowledge congestion Proof Indigestion.
He used a traffic metaphor: in the early 20th century, the existing streets were designed for pedestrians and carriages, and automobiles were suddenly invented and poured directly onto the same streets, resulting in total congestion and paralysis. What this metaphor points to is not that automobiles are bad, but that the road system was not upgraded in sync.
All facts in this article come from a single report by QbitAI (article_id 20655) and are secondhand relayed: neither the original text of Tao's SAIR talk nor the original text of OpenAI's official statement is directly given in the source, and the specific date of the talk and the full name of the SAIR conference are not made clear. Content marked in the article as source_claim—including Villani's reaction, the joint letter of 25 Fields Medalists, the founding and duties of OpenAI's advisory group, and the comment-section netizen reactions—all lack first-party source corroboration and should not be regarded as confirmed facts. Tao's statements about the five-stage lifecycle, AI accelerating only the first two stages, Proof Indigestion, and the risks of creativity loss and data pollution are presented in the source in the form of his own views; this article relays them according to the source's framing and has not independently verified them.
Primary report
Primary source