Terence Tao Warns AI Is 'Depleting' Mathematics — and What That Means for Research

Terence Tao is not a Luddite. The Fields Medalist and UCLA professor has experimented with AI assistants, collaborated on formalization projects, and generally treats new tools as objects of curiosity rather than fear.

That is why his recent warnings land heavily.

In a series of posts on Mathstodon and in interviews with outlets including New Scientist, Tao argues that AI companies are harming mathematics by solving open problems faster than the field can absorb them — dumping proofs without exposition, attribution, or pathways into textbooks and teaching.

He compares indiscriminate automated problem solving to strip-mining a non-renewable resource: the immediate prize is a theorem; the long-term cost is the ecosystem of techniques, failed attempts, and follow-up questions that human exploration would have generated.

For students, educators, and engineers who rely on mathematical progress, the debate is not abstract. It defines how AI should participate in discovery.

Open problems as fuel

Mathematics advances by identifying good problems — questions hard enough to require new ideas, but structured enough to guide a field for years.

Tao writes that AI has flattened difficulty landscapes in many areas. Problems that once signaled depth become benchmarks. Once a solution hits arXiv — or worse, a corporate blog — the problem is "contaminated" for evaluating whether an AI system truly reasoned or recalled training data.

Worse, rumors that someone is working on a problem can trigger preemptive AI sweeps that kill the problem before human researchers finish the slower, richer process of exploration.

The inverse Galois example

Tao co-organized an inverse Galois challenge where contestants found polynomials realizing specific Galois groups. The project deliberately staged competition first (withheld polynomials) to map difficulty, then collaborative release.

Releasing everything at once permanently degraded some problems as training benchmarks — a tradeoff organizers accepted, but not lightly.

That case study generalizes: when speed beats process, you gain answers and lose cartography of the idea space.

Five hurdles before AI math is useful

In New Scientist, Tao outlined five stages any mathematical contribution must pass:

1. Creation — the result exists.
2. Explanation — experts understand the proof.
3. Acceptance — community verification completes.
4. Teaching — material enters curricula.
5. Digestion — the field internalizes techniques as standard tools.

AI companies optimize for stage one. Stages two through five — lower prestige, mostly human labor — fall to academic volunteers Tao describes as demoralized.

"This is not how science is supposed to work," he said. Results may eventually reach textbooks, "but it is just a completely sordid way to get there."

Cultural risk: research goes underground

If sharing a promising direction invites instant AI completion, rational incentives push mathematicians toward secrecy — the opposite of centuries of open science.

Tao fears a future where junior researchers hoard questions, senior mathematicians avoid public brainstorming, and conferences become performance rather than collaboration.

He analogizes to movie spoilers: society enforces norms against ruining narrative experiences. Mathematics may need similar norms — protected problem classes where automated solvers are discouraged.

Hugo Duminil-Copin, another prominent mathematician, uses a different metaphor: great conjectures are lighthouses, not benchmarks for the next model release.

What Tao recommends

His proposals mix ethics and engineering:

• Disclose AI assistance transparently, including chat logs when possible, so reviewers assess reasoning paths — not just outputs.
• De-emphasize "first proof" races; reward clear exposition and teaching.
• Gate publication on whether authors can deliver expert-level, attributed talks — a human digestibility test.
• Declare some problem classes off-limits to automated solvers while directing AI toward domains where rapid solving helps society (industrial optimization, verified code, etc.).

Enforcement is hard. Tao acknowledges that. But social pressure already shapes behavior in academia; he believes protected domains are feasible if paired with sanctioned domains where AI acceleration is welcome.

Lessons for developers and ML engineers

Even if you do not publish theorems, Tao's frame applies to software and ML research:

• Benchmark contamination is real. Once solutions leak into training corpora, evaluation metrics lie.
• Process data — failed attempts, intermediate lemmas — often matters more than final answers for teaching the next generation.
• Speed without documentation creates technical debt at civilization scale.

If you fine-tune models on math corpora or build automated theorem provers, ask: are you contributing to digestion, or only to leaderboard spikes?

A balanced takeaway

Tao is not calling for banning AI in mathematics. He uses AI tools. He wants sustainable extraction rates from the problem supply — and corporate behavior that funds exposition, not just discovery press releases.

AI can multiply human curiosity. It can also consume the curiosity market by answering questions before humans learn to ask them well.

The mathematics community is early in negotiating that boundary. How it resolves will echo through CS theory, physics, and any field where open problems drive progress.

Tao's warning is timely: the machines got fast. The culture has not caught up.

Comments

Loading comments…