Terence Tao's U-Turn: Why AI Math Benchmarks Risk 'Proof Indigestion'

From AI Advocate to Critic

Terence Tao, a Fields Medalist and early adopter of AI in research, previously championed AI tools for literature search, coding, diagramming, and calculation. In March 2026 at the IPAM conference he declared AI "ready for primetime" in mathematics and theoretical physics, calling the shift a "Copernican revolution." Before the 2025 Lunar New Year he envisioned a future where even middle‑school students could contribute to frontier mathematics using AI and Lean, the interactive theorem prover.

The Pivot: "No Reason to Go This Fast"

Half a year later, at a SAIR talk, Tao directly urged model companies to slow down: "AI companies are accelerating endlessly yet know nothing about what comes after. There is absolutely no reason to go this fast." Fellow Fields Medalist Cédric Villani, who had dismissed LLMs as mere statistical parrots three months earlier, said seeing OpenAI claim a Millennium Prize solution felt like "the end of the world."

The Core Problem: Proof Indigestion

Tao's concern is not whether AI can solve problems. He uses a lighthouse analogy: unsolved problems are lighthouses; their value lies in illuminating the surrounding seascape, not the point itself. He outlines a five‑step proof lifecycle:

Generate answer

Lean formal verification

Understand

Digest

Write into textbooks

AI accelerates the first two steps while the last three remain near zero. No reports, no papers, and even the person who presses "enter" cannot explain what just happened. Tao calls this "Proof Indigestion" — a knowledge traffic jam comparable to early 20th‑century streets designed for pedestrians and horse‑carts suddenly flooded with automobiles.

AI Can Verify, But Cannot Produce Insight

One might argue that human mathematicians can still do research while AI churns out proofs. Tao counters that insight cannot be verified, scored, or turned into a training objective. The scorable steps are accelerated; the unscorable ones are left behind. He fears two consequences:

Creativity's source is severed. New problems and directions often emerge from writing textbooks or from serendipitous discoveries during long exploration.

The meandering path of exploration is cut off. AI jumps straight to the final answer, bypassing the wandering process that generates new questions.

A Deeper Hazard Than Data Pollution

Current AI mathematics models are strong because they trained on centuries of human‑digested, textbook‑quality proofs. If the literature becomes flooded with AI‑generated proofs that no human understands, the next generation of models will train on that opaque output. Tao invokes a biological analogy: serial cloning of mice leads to telomere shortening, epigenetic loss, and mutation accumulation — eventually the line dies out. Dumping unverified proofs is, in his view, irresponsible.

Broader Implications: The "Three Questions" from Netizens

The open letter signed by Tao, Peter Scholze, Yitang Zhang, Pierre Deligne, and 21 other Fields Medalists criticizes model companies for treating mathematics as a benchmark to be gamed. OpenAI responded by forming a Mathematics and AI Advisory Group at the Institute for Advanced Study, Princeton, including Edward Witten, Timothy Gowers, and Martin Hairer — but with a disclaimer that the group "is not responsible for advising on internal mathematical progress," effectively welcoming publicity but rejecting oversight.

Online comments revealed a wider anxiety. A top‑voted comment asked: "When painters fell, where was Tao? When writers fell, where was Tao? When programmers fell, where was he?" The article argues that mathematics is merely the first field to confront a shift in value judgment: from "Can AI produce it?" to "Has any human truly digested it?" If the digestion step is skipped, the result is merely a pile of answers no one understands.