Terence Tao argues AI could create 'industrial mathematics' by dividing research labor

Terence Tao outlined a concrete scenario in which AI enables a division of labor long absent from mathematical research: Instead of single researchers framing problems, devising strategies, computing, proving, and writing up results alone, specialized teams would split those tasks and combine generative models with rigorous verification. The shift matters because it could let groups scale ambition and throughput in ways that individual, end‑to‑end craft cannot, effectively creating an “industrial mathematics” model for some lines of inquiry.

Tao identifies two technical levers that must work together: generative AI to propose strategies and automation for formal verification to check proofs rigorously. He warns that if generation outpaces verification, the field will be swamped with plausible‑looking but unvetted ideas. As he puts it, “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.

For the model to succeed, multiple automation capabilities must advance in parallel: idea generation, mechanized proof construction, and formal checking. Tao argues AI could handle routine or computation‑heavy steps inside collaborations — crunching vast numbers of cases or exploring large parameter spaces — while human researchers retain responsibility for high‑level insight, intuition, and the few inspired guesses that guide direction. He envisions teams where AI sifts billions of data points and humans select and refine the most promising leads.

That division of labor carries direct practical implications for builders and research groups. Workflows would shift away from a single researcher’s end‑to‑end craft toward integrated pipelines that combine generative models with verification tooling. Engineering priorities would change accordingly: projects must invest in tight verification and validation chains to prevent low‑quality outputs from proliferating, and teams should design coordination across multiple partial automation components rather than betting on a single, monolithic automated breakthrough.

Tao also stresses a broader lesson about uneven AI performance: human judgment and selective oversight remain indispensable. He frames the challenge as a trade‑off between how much automation is profitable and how stringent verification must be, and he suggests this balance applies beyond mathematics. For builders, the takeaway is that scaling automation responsibly will require parallel progress on verification and on processes that let humans steer, evaluate, and prune AI‑generated proposals.