Terence Tao: AI Could Bring Division of Labor to Mathematics for the First Time
Agentic AIMathematics has never had division of labor. Every mathematician — even the most collaborative — handles the full stack alone: problem formulation, strategy selection, proof execution, verification, and publication. Terence Tao argues AI changes that, for the first time in the field’s history.
The Industrial Mathematics Argument
Speaking at an OpenAI event with Mark Chen focused on AI and mathematical discovery, Tao described a near-term future of “industrial mathematics” — large AI-supported teams pursuing broader research with specialized human and AI roles, mirroring how engineering, medicine, and law already function. Rather than individual researchers working in isolation for years on narrow problems, Tao envisions AI handling what he calls the processing of “billions of data points” while humans contribute “inspired guesses” from limited observations.
The mechanism is formal verification: AI-generated mathematical strategies become useful only when they can be verified rigorously, and verification is now increasingly automatable through proof checkers. Without stringent verification, Tao warns, the output is “slop” — a term he uses precisely, not hyperbolically.
The Critical Constraint
Tao’s optimism is conditional on a single variable: verification quality. His exact framing deserves attention: “The level of automation and AI power that you can profitably use before it becomes slop is roughly proportionate to how stringent your verification is.” This is not a general endorsement of AI-generated mathematics — it is a precise constraint on the conditions under which AI-assisted math remains trustworthy.
The implication for AI in technical and analytical work extends beyond mathematics. The principle — automation’s profitability is bounded by verification stringency — applies equally to AI-generated code, AI-generated clinical decisions, and AI-generated legal analysis. In every domain, the question is not “how much AI?” but “how rigorous is the verification layer?”
The So What
Tao’s framework is the most useful framing yet for thinking about where AI creates genuine capability gains versus where it produces confident-sounding output that nobody can actually validate. The practical test for any AI-assisted analytical workflow is not “did the AI produce something?” but “do we have the verification infrastructure to catch the errors?” Mathematics is the stress test because the standards for correctness are absolute. Every other domain that adopts AI at scale will eventually face the same requirement.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →