AI Is Solving Math's Best Problems Faster Than They Can Be Replaced, Terence Tao Warns

Terence Tao says a new pattern is emerging in mathematical research: powerful AI systems can quickly flatten difficult problems the moment computational effort is concentrated on them, leaving fewer substantive open questions for human researchers to pursue. Tao, a UCLA professor and 2006 Fields Medalist, posted the warning on Mathstodon, arguing that while anyone can pose countless formal questions, only a small subset genuinely advances the field — and those are the ones at risk.

He cited concrete examples of competition between major AI labs. In May, an OpenAI model produced a disproof of the Erdős unit-distance conjecture that outside mathematicians, including Tim Gowers, subsequently verified. In the same week Anthropic ran the same problem through Claude Mythos; Anthropic engineers described Mythos’s output as a shorter proof, and other researchers weighed in with mixed assessments. More recently, Anthropic reportedly formalized a centuries-old proof of Fermat’s last theorem, and OpenAI solved a 90-year-old problem mere hours after a researcher published his own proof, even coauthoring a paper with an Anthropic researcher.

Tao frames these episodes as more than isolated breakthroughs: he says they reveal how an AI-driven race can collapse a problem’s "difficulty landscape" before a human-led project can mature, reversing a long pattern in which new techniques both solved problems and opened fresh directions. Unlike past methodological advances, he argues, current models have uncertain and rapidly shifting capabilities, making it hard to know where their reach ends and which problems still meaningfully test human understanding.

As a practical response, Tao proposed labeling certain questions "analysis-required," so that a bare correct answer produced by a model would be insufficient unless accompanied by reasoning that illuminates related problems. He compared the idea to food banks refusing donations that are merely edible, insisting on contributions that sustain the broader ecosystem. Tao also said that an outright ban on using AI in mathematics is "technically infeasible." His proposal has not been adopted as policy anywhere and may be difficult to implement given how the leading AI labs are already operating.