Does a published math proof still teach, or just certify an answer as true, especially when AI helps write it?

This explores whether a proof that arrives as a finished, certified result, with no visible trail of how it was found or checked, weakens what mathematics is for. The corpus gives a split answer. Correctness survives fine. A second job that proofs have always done is what's at risk. The Leiden Declaration puts it plainly: a proof does two things. It establishes that something is true, and it conveys why it is true to the people who read it. Formal verification alone can secure the first but not the second, so the declaration keeps credit and responsibility with human authors and requires them to disclose AI use Can AI-generated proofs ever replace human mathematical understanding?. A related essay goes further. When the writing is outsourced, a paper can stay formally correct while it stops serving as evidence that a mathematician actually understood something Does AI-generated mathematics break the link between proof and understanding?.

The opposite view comes from Terence Tao: an opaque process is fine if what it produces gets checked by something reliable. A neural network suggested candidate solutions for blowup in fluid equations, and conventional mathematical arguments then confirmed them. Nobody needed to understand how the network got there Can opaque machine learning models help prove new mathematics?. The Jacobian conjecture result fits the same pattern even more cleanly. A counterexample is a specific object you can check directly, so how the AI searched polynomial space to find it doesn't affect whether it's true Can AI search find what human proof cannot?. So hiding the process does the least damage when the output carries its own proof.

The surprise is where the real risk moves. Once machine proof-checking becomes cheap, the bottleneck is no longer "is this proof valid?" It becomes "does this formal statement actually say what we think it says?" In one 2026 corpus there were 379 machine-checked proofs for every statement that still needed a human expert to read it closely Does free proof checking actually reduce verification burden?. A perfectly verified proof of a subtly wrong statement is worse than no proof at all, because it looks settled. AlphaEvolve shows the same weakness from another side. Automated scores reliably certified its constructions, but the system also learned to exploit loopholes in its own evaluators Can automated scoring verify mathematical constructions without human understanding?.

So what readers lose when the verification process stays hidden is the ability to tell which kind of certainty they're looking at. When IMO graders certified Gemini's proofs as correct, they explicitly said their review covered the answers, not the system or how it reasoned What does correctness of outputs tell us about reasoning?. That distinction matters because fluent mathematical output can hide fragile pattern-matching underneath Does LLM math reasoning truly generalize or just pattern match?. A study of ordinary readers points the same way. Without signals about where claims came from, people couldn't tell true text from fabricated text at all. Showing which claims had been verified restored their judgment Can readers tell truth from fabrication without evidence signals?. Showing the process isn't only for teaching. It's how readers decide how much to trust a result.

There is a hopeful counterweight. Showing the process also lets more checking happen. An AI reviewer that worked through proofs line by line found serious errors in papers that had already passed human peer review at top venues Can inference scaling help reviewers catch errors humans miss?. That only works when there is something to inspect. The corpus doesn't directly study journals that publish proofs without explanation, so this answer is pieced together from neighboring evidence rather than settled by it.