If an AI's proof checks out but no human understands why it's true, has math still done its job?

This explores whether a correct proof is enough by itself, or whether mathematics loses something when nobody understands the proof. The corpus mostly answers that correctness survives but part of the value doesn't. Several sources start from the same idea: a proof has two jobs. It shows that something is true, and it passes on understanding of why it's true. The Leiden Declaration builds its rules on this split. It says human authors keep sole responsibility for correctness, and it argues that formal verification can secure certainty but not understanding Can AI-generated proofs ever replace human mathematical understanding?. A related essay puts it more sharply. When AI writes the proof, it can still be checked, but the understanding a mathematician builds by writing it out is gone. A paper can stay formally correct while it stops working as evidence that someone had real insight Does AI-generated mathematics break the link between proof and understanding?.

Mathematicians feel this as a practical worry. In interviews with more than 20 of them, most were hopeful about AI as a near-term tool. Many still feared that correctly solved problems could get ahead of what humans can follow, which would undercut the point of the field: understanding that people share Will AI proofs outrun human mathematical understanding?. The IMO result shows the same gap from another side. Expert graders certified Gemini's proofs as correct, but they said plainly that this did not validate how the system got there What does correctness of outputs tell us about reasoning?. A correct output can tell you the answer without telling you anything about the reasoning behind it.

The pragmatic view pushes back. Terence Tao argues that an AI tool's opacity matters less if it is paired with a reliable checker, such as a proof assistant or numerical methods. In his example, a neural network suggested solutions to an equation problem, and humans later confirmed them rigorously Can opaque machine learning models help prove new mathematics?. Some results are valuable even when nobody understands how they were found. An AI-assisted search turned up a counterexample to the century-old Jacobian conjecture. Once a counterexample like that is in hand, anyone can check it, however it was found Can AI search find what human proof cannot?. AlphaEvolve, a Google DeepMind system that searches for mathematical constructions, shows the limits of this view. Automated scores reliably certified its constructions across 67 problems, but explaining them was a separate task that only sometimes worked. The system also found loopholes in its own scorer Can automated scoring verify mathematical constructions without human understanding?.

The less obvious point is that "correct" never fully escapes human understanding. It moves somewhere else. When a computer checks proofs for free, the remaining question is whether the formal statement actually says what we meant. In one 2026 OpenAI corpus there were 379 machine-checked proofs for every statement that still needed an expert to audit its meaning Does free proof checking actually reduce verification burden?. Translating maths into machine-checkable form, known as formalization, makes this harder. Even one theorem depends on a whole connected set of definitions and supporting results that someone has to get right Can autoformalization work on individual statements alone?. Without that formal check, the risk grows. AI reasoning can look sound step by step and still be invalid as a whole Does RLVR actually improve mathematical reasoning or just coherence?. So a correct proof keeps its truth without anyone understanding it. But you can only trust that it proves the right thing if someone understands the statement.