Terence Tao, a Fields medalist, warns that AI systems are approaching the ability to solve real research-level mathematics problems cheaply and at scale—a shift he compares to the foundational crisis that shook mathematics between 1900 and 1930.
Rather than questioning what AI can compute, Tao's concern is deeper: if AI-generated proofs flood the field faster than humans can verify, understand, or learn from them, mathematics risks losing its core values around community, training, and genuine understanding.
He proposes that results should only be published if authors can convincingly explain them in expert-level talks.
What happened
Fields medalist Terence Tao argues that AI systems will soon be capable of performing a reasonable fraction of research-level mathematical tasks with reasonable success and cost. In the First-Proof Project's second round, seven of ten never-published research problems received passing grades from at least one of four AI systems tested, with solutions costing tens to hundreds of dollars per problem.
Why it matters
Tao sees this as a foundational crisis comparable to the 1900–1930 crisis triggered by Russell's paradox and Gödel's incompleteness theorems. The shift threatens to upend not mathematical truth but the implicit framework of what counts as a contribution, what gets rewarded, and what it means to understand something. If AI floods the field with cheap proofs faster than humans can verify or learn from them, the discipline risks fragmenting core goals—solving problems, building theories, fostering community, and training the next generation—that have always reinforced each other.
What to watch
Tao proposes a practical guardrail: results should not be published unless authors can demonstrate they can give a clear, expert-level talk on their findings that is correct and properly attributed. He cites the Leiden Declaration on Artificial Intelligence and Mathematics (published June 2026, backed by the International Mathematical Union) as formal guidance. The Erdős problem database already contains dozens of AI-generated submissions awaiting human expert verification.
Ask the AI about this article →
Terence Tao's concern is not that AI will fail at mathematics, but that it will succeed in ways that fragment the discipline's core identity. The First-Proof Project results show AI systems are already solving hard, unpublished research problems at scale and cost that undercut the traditional scarcity model. This abundance, Tao argues, exposes a tension long masked by resource constraints: mathematics has always woven together multiple goals—proof-finding, theory-building, community formation, and mentorship—that reinforce each other. AI, by excelling at the measurable and quotable, threatens to pull them apart.
Tao's diagnosis rests on Goodhart's law: when a measure becomes a target, it ceases to be a good measure. Generative AI is structurally prone to this because it optimizes for the appearance of a good output rather than the thing itself. The AI industry's financial incentives make it worse, rewarding exactly the benchmarkable wins that mathematicians have long used as proxies for deeper achievement. If AI-generated proofs pile up faster than humans can check or absorb them, the field risks substituting abundance for understanding—a crisis not of truth but of meaning.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
