Terence Tao: Mathematics in the Age of AI
Bilibili video BV11T8A6FEgC. The page exposes no public subtitle track, so this is an analysis of the lecture topic, Tao’s essay, and the public slides, not a verbatim transcript.
The central question
The interesting question is not simply whether AI can produce a proof. Mathematical research is a pipeline: choose a valuable problem, explore examples, form a conjecture, produce a candidate argument, verify it, explain it, and let a community reuse it.
AI can suggest lemmas, translate notation, search large corpora, and generate Lean code. A proof assistant can reject an invalid formal term. Neither one decides whether a result is important, understandable, or worth teaching.
Abundance changes the bottleneck
Ten candidate proofs can be read. Ten thousand candidates create a problem of attention, ranking, verification, and explanation. More theorems are not automatically more progress: mathematics also values theory, understanding, education, collaboration, and durable exposition.
This is close to Goodhart’s law. If publication counts or benchmark scores become the target, systems may optimize proof-shaped output rather than useful mathematics. The danger is not that correct results become false; it is that the stream becomes faster than the community’s ability to judge it.
Formal verification and human judgment
Lean can form a precise interface between human ideas and machine generation. The prover checks a narrow obligation; people choose definitions, strategies, and explanations. Formalization has costs, and a checked proof may still be difficult to understand.
The likely workflow is layered: informal exploration for speed, formal checking for reliability, and human exposition for transfer. Tool use should be documented as provenance: readers need to know what was generated, checked, and rewritten.
Conclusion
Tao’s lecture is less a prediction that mathematicians disappear than a warning that the community must decide what it values before automation decides through sheer volume. AI can generate candidates and proof assistants can check steps. Humans still decide what deserves understanding, explanation, and transmission.