Mathematicians warn AI is closing in on research-grade problem solving

Mathematicians are growing uneasy as large language models move from struggling with basic arithmetic to tackling problems that resemble genuine research mathematics. Recent benchmarks and informal tests suggest systems are no longer just pattern-matching textbook exercises but producing proof sketches and solutions that require real conceptual work, narrowing what was assumed to be a comfortable gap between machine output and expert human reasoning.

The concern is twofold. Practically, the field has to reckon with verification: AI-generated proofs can look plausible while hiding subtle errors, putting more burden on human reviewers and on formal proof systems like Lean. Culturally, mathematicians are debating what their role becomes if machines can routinely propose, and eventually verify, novel results — and how training, credit, and funding should adapt before the shift outpaces the community’s ability to respond.