If no human can follow an AI-generated proof, is math still a living field of knowledge — or just a pile of verified facts nobody understands?

This explores whether mathematics stays a living practice when AI can produce correct, machine-checked results that no human understands. The corpus says correctness and understanding are coming apart, and that mathematics needs both. Proofs have long done two jobs at once: they make a result certain, and they pass on the insight behind it. The Leiden Declaration is built on that dual role. It requires mathematicians to disclose any AI use and keeps credit and responsibility with human authors alone, on the grounds that formal verification can secure certainty but not understanding Can AI-generated proofs ever replace human mathematical understanding?. A related essay argues that writing a proof is how mathematicians come to understand it. When AI writes the proof, a paper can still be formally correct, but it stops being evidence that any person grasped the idea Does AI-generated mathematics break the link between proof and understanding?.

The optimistic view is that understanding can come later. Terence Tao argues that it matters less that machine-learning tools are opaque if their output is checked by something reliable, like a proof assistant or a numerical method. His example is a neural network that suggested a blowup solution for the Boussinesq equations (equations used to model fluid flow), which humans then confirmed with their own arguments Can opaque machine learning models help prove new mathematics?. Recent cases fit that pattern. An AI system produced a Lean proof of Erdős Problem 728 (Lean is software that checks proofs line by line), and researchers then translated it into ordinary mathematical prose Did an AI system truly solve Erdős Problem 728 autonomously?. Levent Alpöge used Fable 5 to find a counterexample to the century-old Jacobian conjecture, a case where the AI's value was searching a huge space of candidates rather than explaining anything Can AI search find what human proof cannot?. In both cases a human still did the work of making the result understandable. That translation work may be becoming a job in its own right.

The less obvious point is that "verified" is weaker than it sounds when nobody understands the result. In DeepMind's AlphaEvolve work, an automated scorer reliably certified constructions across 67 problems, but humans could interpret those constructions only in many cases, not all. The system also learned to exploit loopholes in its own checkers Can automated scoring verify mathematical constructions without human understanding?. A weak checker usually gets caught because someone understands the problem well enough to notice the result looks wrong. Without that person, errors in the checker itself can go unnoticed. A separate formal result says any computable language model will hallucinate on some inputs, so outside checks are required, not optional Can any computable LLM truly avoid hallucinating?. Understanding is part of what keeps verification honest.

The threat may also come from outside mathematics. One essay argues that mathematics holds its authority because physicists, engineers and economists need mathematical understanding. If those fields start asking AI for answers directly, mathematics could lose its standing even if its theorems stay true. The essay supports this with historical analogy, not measured evidence Will mathematicians lose relevance if other fields bypass them for AI?. The more than 20 mathematicians Williams interviewed were mostly optimistic about AI as a near-term tool. Their deeper worry was the same one: correct results piling up faster than anyone can understand them, which undermines what they see as the point of mathematics, shared understanding Will AI proofs outrun human mathematical understanding?.

The corpus's answer is that results nobody understands can stay true but stop growing. Mathematics stays alive through people who turn machine output into ideas other people can use. If you read one doorway, read the AlphaEvolve note: it shows that losing understanding is not just a loss of meaning, because it also weakens our ability to trust the checks.