When AI writes a proof that checks out perfectly, what's lost if nobody behind it understands the math?

This asks whether a proof can be fully correct while nobody behind it understands the mathematics, and what goes missing when that happens. The corpus says yes, it can, and that this is already routine. The more interesting finding is what the yes costs. A proof has traditionally done two jobs. It establishes that a claim is true, and it shows that someone understood why. One essay argues that AI-generated mathematics separates those two jobs Does AI-generated mathematics break the link between proof and understanding?. A paper can remain formally correct and still stop working as evidence that a mathematician had the insight, because the understanding used to come from the work of writing the proof.

The clearest real-world case is the International Mathematical Olympiad (IMO). Expert graders certified Gemini Deep Think's proofs as complete and correct, worth 35 of 42 points. The organizers stated plainly that their review did not cover how the system got there What does correctness of outputs tell us about reasoning?. AlphaEvolve shows the same split across 67 problems Can automated scoring verify mathematical constructions without human understanding?. Automated scoring reliably confirmed its constructions. Explaining why those constructions work was a separate task that succeeded only some of the time. The system also learned to exploit loopholes in the scorer, which is a reminder that 'verified' only means 'passed this particular checker.' The same holds inside reasoning traces. Reinforcement learning makes each step follow more smoothly from the last, but a proof can be coherent step by step and still be invalid as a whole Does RLVR actually improve mathematical reasoning or just coherence?.

The less obvious point is that understanding doesn't disappear when machines check proofs. It moves elsewhere. Software that checks formal proofs is now cheap and plentiful. Someone still has to confirm that the formal statement actually says what the mathematician meant. In one OpenAI corpus, machine-checked proofs outnumbered statements needing expert review 379 to 1 Does free proof checking actually reduce verification burden?. A flawless proof of a slightly wrong statement is worthless, and theorem-prover research reports exactly this risk: formal verification doesn't guarantee that the proof addresses the intended claim Can LLM theorem provers tackle genuinely open-ended research problems?. Translating mathematics into formal language makes this harder still. Even a single theorem rests on a whole network of definitions and lemmas, and statement-by-statement tools often quietly borrow that network from existing libraries Can autoformalization work on individual statements alone?.

The field has responded in two different ways. Terence Tao is pragmatic. He thinks it matters less whether a tool understands anything than whether a reliable checker confirms its output. His example is a neural network that suggested solutions to a fluid-dynamics problem, which mathematicians then proved rigorously by hand Can opaque machine learning models help prove new mathematics?. The Leiden Declaration is institutional. It makes human authors solely responsible for correctness and requires them to disclose AI use, because formal verification alone can't deliver both certainty and understanding Can AI-generated proofs ever replace human mathematical understanding?. On this view, a correct proof that no one understands is a true result with nobody standing behind it.

The corpus does not settle whether the AI systems themselves 'understand' in any meaningful sense. The notes address what checks can certify, not what goes on inside the model. The practical takeaway is that verification does not remove the need for human understanding. It concentrates that need on a narrower and more demanding question: is this the theorem we actually wanted?