How Fields Medalist Terence Tao is using AI to reshape the boundaries of mathematical proof

For decades, pure mathematics remained the ultimate redoubt of un-digitized human genius—an art form of pure logic too subtle for computers to do anything but calculate. But according to Terence Tao, the world’s most prominent living mathematician, that era is officially over. In a newly published slide deck prepared for the International Congress of Mathematicians (ICM) 2026, the Fields Medalist outlines a stark future where AI in mathematics is no longer an experimental curiosity, but the standard operating system for proving the impossible.

The Convergence of Intuition and Rigor

To understand why Terence Tao’s endorsement of AI in mathematics matters, one must look at his recent obsession with formal proof assistants. For the past few years, Tao has been publicly experimenting with Lean, an open-source theorem prover and functional programming language developed by Leonardo de Moura at Microsoft Research. The slide deck, titled "Mathematics in the Age of AI," signals that what began as an eccentric hobby for some of the world's top minds is now the official frontier of the discipline.

Historically, mathematical discovery has been split into two phases: the messy, intuitive leap to generate a conjecture, and the grueling, bureaucratic process of formalizing the proof. Traditionally, humans did both. But as proofs have ballooned into hundreds of pages of hyper-dense abstractions, peer review has begun to buckle under its own weight. By leveraging automated theorem provers alongside large language models (LLMs), mathematicians are outsourcing the bureaucracy while supercharging the intuition.

Why Mathematics Resisted AI (And Why That Just Changed)

For years, traditional machine learning models struggled with mathematics. Deep learning is probabilistic; it operates on statistical patterns, predicting the next most likely token. Pure mathematics, however, is binary and unforgiving: a proof is either 100% correct, or it is useless. A single hallucinated decimal or false assumption invalidates the entire structure.

The breakthrough identified by Tao and other researchers is the marriage of neural networks with symbolic logic engines. Large language models excel at generating plausible mathematical tactics, suggesting leaps of intuition, and translating natural human reasoning into formal code. The formal proof assistant (like Lean or Coq) acts as the ultimate compiler, instantly checking the LLM's output for mathematical validity. This combination of "neural" intuition and "symbolic" rigor effectively eliminates the hallucination problem, transforming LLMs from unreliable copywriters into hyper-efficient brainstorming partners.

"The integration of formal proof assistants and neural networks creates a self-correcting loop. The AI suggests the path, but the math compiler enforces the truth."

Terence Tao, ICM 2026 Slides

The Rise of the Mathematical "Copilot"

What does this look like in practice? Tao’s slides suggest that the workflow of the modern mathematician is shifting toward that of a software engineer. Instead of writing proofs on blackboards, researchers write code in Lean. The AI acts as a sophisticated autocomplete engine, filling in trivial lemmas, searching massive mathematical libraries like Mathlib, and pointing out logical gaps in real-time.

This dynamic was famously demonstrated when Tao used Lean to formalize a proof of the Polynomial Freiman-Ruzsa conjecture. What would have taken months of meticulous human cross-referencing was completed in a fraction of the time, with absolute mathematical certainty. This is not just about speed; it is about cognitive scale. By automating the verification of intermediate steps, mathematicians can attempt vastly more complex structures without fearing that a foundational error fifty pages back will ruin their life's work.

The Feedback Loop: Why Frontier AI Needs Math

While mathematicians are eager to use AI, the AI industry is equally desperate for mathematics. As frontier AI labs like OpenAI, Anthropic, and Google DeepMind run out of high-quality public internet data to train their models, mathematical proofs offer a goldmine of clean, structured synthetic training data.

AI models trained on formalized mathematics do not just get better at math; they get better at reasoning. The rigid logic required to satisfy a proof assistant translates directly to improved planning, coding, and multi-step problem-solving capabilities in LLMs. Tao's work represents the bridge: as mathematicians formalize the world's mathematical knowledge graph, they are building the ultimate dataset for the next generation of reasoning-centric artificial intelligence.

The Takeaway

Terence Tao's ICM 2026 presentation is a clear signal to the global scientific community: the line between human and machine intelligence is no longer a barrier, but an interface. The future of mathematics will not belong to silicon alone, nor will it remain the exclusive domain of the human brain. It belongs to the hybrid researcher who can orchestrate both.