Terence Tao Charts Mathematics' Future in the Age of AI

For most of its history, mathematics has been the last discipline anyone expected a machine to master. Arithmetic could be automated centuries ago, but the act of proving something new — of arriving at a chain of reasoning that no one had ever assembled before — seemed to belong permanently to human minds. That assumption is now under open review. On July 24, 2026, at the International Congress of Mathematicians, the field's most celebrated living practitioner stood before his peers and argued that the era of AI-assisted proof has quietly, and irreversibly, begun.

What Happened at the ICM 2026 Stage

Fields Medalist Terence Tao delivered a public lecture titled "Mathematics in the Age of AI," a talk that functioned less as a product demonstration than as a map of a discipline in transition. Rather than asking whether artificial intelligence can do mathematics, Tao broke the work of a mathematician into distinct stages that extend well beyond the romantic image of a lone thinker cracking a problem. There is the generation of a proof, then its verification, its exposition so that others can follow it, its publication, and finally its canonicalization — the slow process by which a result is absorbed into the shared body of knowledge that future generations inherit.

His central claim was that AI is no longer confined to the margins of these stages. Modern systems are already strong at generating candidate proofs and are advancing quickly at verification. To ground the point, Tao cited an experiment he called "First Proof," run on May 28, in which four AI harnesses were set against ten novel, research-level mathematics problems under controlled conditions. For seven of the ten, at least one system produced a solution judged to be of publication quality by expert referees, who assessed both correctness and clarity of exposition. The compute cost ranged from roughly ten to a thousand dollars per problem — an accounting detail that, more than any rhetorical flourish, signals how ordinary this capability is becoming.

What makes the framing striking is its restraint. Tao did not announce that mathematics had been solved or that human mathematicians had been rendered obsolete. He described, in careful terms, a workflow already shifting beneath the profession's feet, and invited the audience to think about what a discipline does when the scarcest resource — a correct, novel proof — starts to become abundant.

Why It Matters Beyond Pure Mathematics

To understand why a slide deck about theorem proving drew attention far outside academic mathematics, it helps to recall why the subject resisted machine learning for so long. Deep learning is fundamentally probabilistic; it predicts the most plausible next token and tolerates a margin of error. Mathematics tolerates none. A proof is either valid or worthless, and a single unjustified step, however small, collapses the entire structure. For years this all-or-nothing character made pure mathematics the natural adversary of systems that occasionally hallucinate.

The change Tao describes comes from pairing two very different technologies. Large language models supply intuition — they propose tactics, suggest which lemma might close a gap, and translate loose human reasoning into formal code. A proof assistant such as Lean, the open-source system created by Leonardo de Moura, then acts as an uncompromising compiler, checking every line against the axioms and refusing anything that does not hold. The language model can guess; the proof assistant enforces the truth. In that loop, the hallucination problem that plagues AI elsewhere is largely neutralized, because nothing reaches the final proof unless the formal checker has certified it.

This is why the talk resonated with engineers and scientists who have never opened a copy of the Mathlib library. The same rigor that satisfies a proof assistant translates into more disciplined reasoning across other domains. A model trained to survive the merciless scrutiny of formal verification tends to plan better, code more carefully, and handle multi-step problems with fewer silent errors. Mathematics, in other words, is turning out to be one of the cleanest training grounds available for the reasoning capabilities that frontier AI labs are racing to improve.

The Reaction from the Mathematical Community

Tao's own credibility gives the argument unusual weight. He is not a technologist selling a platform but a working mathematician who has spent recent years publicly experimenting with formal methods, including a widely discussed effort to formalize a proof of the Polynomial Freiman-Ruzsa conjecture using Lean. What might once have taken months of meticulous cross-referencing was compressed dramatically, with the added guarantee that no hidden error lurked in the intermediate steps. For a community that has watched proofs balloon to hundreds of dense pages, straining the capacity of any human referee, the appeal is not hard to grasp.

Still, the response has been far from uniform enthusiasm. Some mathematicians welcome the prospect of offloading the bureaucratic labor of verification while preserving the creative core of the work for humans. Others worry about what is lost when proofs are generated faster than anyone can absorb them, and about a subtle shift in the meaning of understanding. There is a difference between knowing that a statement is true because a checker confirmed it and understanding why it is true in a way that reshapes intuition. Tao's framing acknowledges this tension rather than dismissing it, which is part of why the talk has circulated so widely.

The debate also touches on questions of access and equity. Formal tools have a steep learning curve, and there is a real risk that a two-tier profession emerges, split between those fluent in the new machinery and those who are not. The optimistic reading is that AI copilots could eventually lower the barrier rather than raise it, filling in routine lemmas and searching vast libraries so that a researcher can focus on the ideas that matter.

What Comes Next for Proof and Trust

The most provocative thread in Tao's argument is that the profession's bottleneck is moving. For centuries the scarce thing was the proof itself — the hard-won argument that a claim is true. If frontier systems can already produce publication-quality solutions to novel problems much of the time, then generating a proof is no longer the constraint. The new scarcity is trust: deciding which of many machine-generated proofs deserves to be believed, taught, and built upon.

This reframing has practical consequences. Verification and canonicalization, the later stages Tao outlined, suddenly become the center of gravity rather than an afterthought. A field that produces more proofs than any human can read will need robust ways to certify them, to compress them into human-legible explanations, and to weave them into the existing knowledge graph without contradiction. Formal libraries like Mathlib, which encode mathematics in machine-checkable form, look less like a niche hobby and more like critical infrastructure for a discipline about to be flooded with results.

There is a feedback loop lurking here as well. As mathematicians formalize more of their knowledge, they generate exactly the kind of clean, structured, unambiguous data that reasoning-focused AI systems crave. The better those systems become, the more mathematics they can help formalize, which in turn produces still more training material. Whether this loop accelerates smoothly or introduces new failure modes is one of the genuinely open questions the talk leaves on the table.

Closing Thoughts on a Discipline in Transition

What lingers after reading Tao's slides is not a sense of triumph or alarm but of a threshold being crossed thoughtfully. He is describing a mathematics that neither belongs to silicon alone nor remains the exclusive province of the human brain, but is instead practiced at the interface between the two. The mathematician of the near future may look more like an orchestrator — someone who poses the right questions, judges which machine-generated arguments are worth trusting, and preserves the human thread of understanding that gives a proof its meaning.

It is worth resisting the temptation to read this as a story about machines replacing people. The more interesting possibility is that abundance changes the questions we ask. When proofs were scarce, ingenuity was spent producing them; when they are plentiful, ingenuity may migrate toward taste, judgment, and the harder problem of deciding what is worth proving in the first place. Tao's lecture is, at heart, an invitation to sit with that shift rather than to celebrate or fear it — a reminder that the deepest changes in a field often arrive not as a single breakthrough but as a quiet redefinition of what the work itself is for.

한글 요약

필즈상 수상자 테렌스 타오가 2026년 7월 24일 국제수학자대회(ICM)에서 "AI 시대의 수학"이라는 공개 강연을 열고, AI가 보조하는 증명의 시대가 이미 조용히 시작됐다고 진단했습니다. 그는 수학 작업을 증명 생성, 검증, 서술, 출판, 그리고 후대에 전승되는 정전화(canonicalization)라는 단계로 나누고, AI가 이미 증명 생성에 강하며 검증에서도 빠르게 발전하고 있다고 설명했습니다. 근거로는 5월 28일 진행된 'First Proof' 실험을 들었는데, 네 개의 AI 시스템을 연구 수준의 신규 문제 10개에 투입한 결과 그중 7개에서 전문가가 출판 가능 수준으로 판정한 해답이 나왔고, 문제당 연산 비용은 약 10~1000달러였습니다.

핵심은 대형 언어모델의 직관과 Lean 같은 형식 증명 보조기의 엄밀함을 결합한 구조입니다. 언어모델이 전략과 보조정리를 제안하면, 형식 검증기가 모든 단계를 공리에 비추어 냉정하게 확인해 오류를 걸러냅니다. 이 되먹임 구조 덕분에 다른 분야에서 문제되는 환각이 수학에서는 크게 억제되며, 동일한 엄밀함이 계획·코딩·다단계 추론 같은 AI의 일반 능력 향상으로도 이어집니다. 타오는 이미 Lean으로 폴리노미얼 프라이만-루자 추측 증명을 형식화한 경험이 있어, 그의 진단은 기술 판매자가 아닌 현역 수학자의 관점에서 무게를 지닙니다.

가장 도발적인 대목은 병목이 이동한다는 통찰입니다. 오랫동안 희소했던 것은 증명 자체였지만, AI가 신규 문제의 출판급 해답을 자주 만들어내는 지금은 '어떤 증명을 신뢰할 것인가'가 새로운 과제가 됩니다. 이에 따라 검증과 정전화, 그리고 Mathlib 같은 기계 검증 가능한 형식 라이브러리가 학문의 핵심 인프라로 부상합니다. 타오의 강연은 승리도 경고도 아닌, 수학이라는 학문이 인간과 기계의 접점에서 다시 정의되는 전환점을 차분히 사유하자는 초대에 가깝습니다. 참고: ICM 2026 슬라이드, VnExpress, Logic of Logic.