Terence Tao Is Wrong About AI Killing Math Because Math Was Already Dead

Hand-wringing over whether artificial intelligence will destroy pure mathematics is the ultimate academic vanity project.

Nature recently ran a piece asking if math can survive AI, spotlighting Fields Medalist Terence Tao’s call for a "Maths 2.0" era. The premise is polite, prestigious, and completely backwards. It assumes that human-driven mathematics is currently a thriving, pristine cathedral of pure thought that might be bulldozed by Silicon Valley’s probabilistic bulldozers.

That cathedral collapsed decades ago. It just took a transformer model to force mathematicians to look at the rubble.

For the past fifty years, high-level mathematics has operated as an artisanal guild masquerading as an objective science. Papers have grown so bloated, hyper-specialized, and unreadable that a breakthrough in analytic number theory can sit unverified for years simply because only four people on Earth possess the niche vocabulary to audit the proof. Peer review in pure math isn't a rigorous verification engine anymore; it is a trust-based gentleman's agreement.

When Shinichi Mochizuki dropped his 500-page proof of the abc conjecture in 2012, the mathematical establishment didn't celebrate. They froze. Over a decade later, the field remains fractured over whether the proof actually works, not because the logic is inherently unknowable, but because human cognitive bandwidth has hit a hard evolutionary wall.

AI isn't invading math to kill it. AI is exposing the dirty secret that human mathematicians ran out of working memory around 1980.

The Lie of the Lone Genius

The romanticized image of mathematics—a solitary thinker with a piece of chalk, conjuring eternal truths from the ether—is the exact delusion holding the discipline back. We fetishize the struggle of the proof rather than the validity of the theorem.

Terence Tao, to his credit, recognizes that formal verification tools like Lean are necessary. He used Lean to formalize the Polynomial Freiman-Ruzsa conjecture, turning a dense human argument into machine-checkable code. Yet the broader academic establishment treats this shift like a tragic compromise, as if using a compiler somehow sullies the purity of the human intellect.

Let's be brutally honest about how modern software engineering works versus how academic math works. If a senior systems engineer pushed 400 pages of uncompiled, untested C++ code to a production repository and asked three colleagues to "read it over and vouch for its correctness," they would be fired before lunch. Yet in pure mathematics, publishing 400 pages of uncompiled prose and calling it an immutable truth is how you win a Fields Medal.

I have watched quantitative research teams burn millions of dollars trying to implement "proven" theorems from peer-reviewed journals, only to discover subtle edge-case flaws buried on page 112 that invalidate the entire model. Why? Because humans are terrible at tracking state across ten thousand logical dependencies. We get tired. We skip steps we consider "trivial." We let intuition paper over rigor.

Formal proof assistants—and the Large Language Models now learning to write code for them—do not get tired. They do not find a lemma "obvious." They demand absolute, uncompromising syntactic truth.

What "Maths 2.0" Actually Means

To understand why the current panic is misplaced, you have to separate two completely different technologies that the media constantly conflates:

  1. Generative LLMs (GPT-4, Claude, Gemini): These are probabilistic token predictors. Left alone, they will hallucinate a plausible-sounding proof of the Riemann Hypothesis that collapses under three seconds of scrutiny. They are intuitive engines, brilliant at pattern matching and terrible at deterministic logic.
  2. Interactive Theorem Provers (Lean, Coq, Isabelle): These are strict deterministic compilers based on dependent type theory. They have zero intuition. They cannot guess. They simply verify whether step B follows legally from step A according to the foundational axioms of mathematics.

The real revolution—the one that renders the "can math survive AI?" question obsolete—happens when you wire the hallucinatory intuition of an LLM directly into the unforgiving cage of a theorem prover.

Think of the LLM as a hyper-caffeinated graduate student throwing wild, creative guesses at a whiteboard, and Lean as a ruthless, incorruptible bouncer checking IDs at the door. The LLM can propose ten thousand nonsensical leaps of logic, and it doesn't matter, because Lean rejects 9,999 of them in milliseconds. The single step that passes the compiler is guaranteed to be mathematically bulletproof.

DeepMind’s AlphaProof and AlphaGeometry already demonstrated this exact architecture at the International Mathematical Olympiad, achieving silver-medal performance on problems that stump the brightest teenagers on the planet. They didn't "ruin" the competition. They proved that search combined with formal verification scales faster than biological evolution.

Dismantling the "Black Box" Panic

Whenever this topic surfaces, the same tired objections emerge from tenured faculty lounges. Let's tear through the most common myths.

"If a machine proves a theorem, humans won't understand why it's true."

This is pure intellectual narcissism. The universe does not owe human beings intuitive comprehension of high-dimensional topology. In 1976, Kenneth Appel and Wolfgang Haken proved the Four Color Theorem using a computer to check 1,936 configurations. Purists screamed that it wasn't "real" math because no human could read the proof in a single sitting. Fifty years later, the Four Color Theorem is still true, and the purists still haven't found a short, elegant human proof. Understanding is a psychological comfort, not a mathematical requirement.

"AI will automate away the need for human mathematicians."

If your definition of a mathematician is "someone who grinds through routine algebraic manipulations to publish incremental papers in low-tier journals," then yes, you are obsolete. Good. The academic publishing industry is drowning in useless, unreadable PDFs designed solely to pad CVs for tenure committees. Automating the grunt work of proof verification frees humans to do the only thing machines still suck at: asking the right questions. A theorem prover can tell you if a statement is true, and an LLM can help find the path to prove it, but neither of them possesses the taste, curiosity, or philosophical drive to decide which conjectures actually matter.

"LLMs are fundamentally bad at math because they can't even multiply large numbers."

This argument died in 2023. Criticizing an LLM for failing at raw arithmetic is like criticizing a Formula 1 driver for being slow on foot. You don't ask a neural network to do mental math; you give it access to a Python interpreter and a formal verification kernel. Tool-use completely neutralizes the deterministic weaknesses of neural networks.

The Ugly Truth About the Transition

Of course, pretending this transition will be painless is just as dishonest as predicting the death of the field. There are massive, structural downsides to turning mathematics into a branch of formal software engineering.

First, the barrier to entry is about to skyrocket in the wrong direction. Learning Lean or Coq is notoriously brutal. Right now, formalizing a single page of textbook mathematics can take an expert dozens of hours of painstaking code translation. If the mathematical establishment mandates formal verification before an AI-assisted proof is accepted, we risk sidelining brilliant intuitive thinkers who simply aren't good programmers. We could easily trade an ivory tower of obscure prose for an ivory tower of unreadable functional code.

Second, compute inequality will corrupt pure math. Historically, pure mathematics was the ultimate egalitarian science. All you needed was a pencil, a ream of paper, and a brain. You didn't need a particle accelerator or a wet lab. If "Maths 2.0" requires massive reinforcement learning runs across thousands of H100 GPUs to search proof trees, then cutting-edge mathematical discovery becomes the exclusive property of frontier AI labs and billion-dollar endowments. A genius working out of a cramped apartment in Eastern Europe or South Asia won't be able to compete with Google DeepMind's compute budget.

Third, we risk optimizing for provability over depth. When you hand researchers a powerful search tool, they naturally gravitate toward problems that the tool is good at solving. Combinatorics and finite geometry yield easily to automated search; deep, conceptual bridges between distant fields—like the Langlands program—do not. There is a very real danger that AI turns mathematics into a high-speed factory for churning out thousands of technically impressive, completely trivial theorems while the foundational mysteries remain untouched.

Stop Mourning the Chalkboard

If you are a researcher, a student, or a technologist watching this unfold, stop reading elegies about the soul of mathematics and adapt to the new stack.

  1. Treat prose proofs as rough drafts. If an argument isn't formalized in a proof assistant, treat it as an uncompiled hypothesis, not an established fact. Learn the basics of Lean 4. It is the new LaTeX. Within a decade, submitting a major paper without a formal verification repository attached will look as amateurish as submitting a handwritten manuscript.
  2. Shift your focus from verification to architecture. Stop competing with machines at algebraic endurance. Focus on problem formulation, cross-domain synthesis, and building the high-level conceptual frameworks that guide automated search. Be the architect, not the bricklayer.
  3. Demand open-source verification libraries. Do not let private tech monopolies own the formalized corpus of human mathematical knowledge. Support open, community-driven initiatives like Lean's mathlib, where every proven theorem becomes a reusable, machine-readable building block for the entire world.

Mathematics isn't dying. It is finally shedding its biological training wheels. The mathematicians who survive won't be the ones who cling to the romantic misery of the chalkboard; they will be the ones ruthless enough to let machines handle the rigor while they hunt for the truth.

LE

Lucas Evans

A trusted voice in digital journalism, Lucas Evans blends analytical rigor with an engaging narrative style to bring important stories to life.