It's almost always easier to understand theorems than proofs. People could be writing down properties of their programs and the AI could generate a proof or produce a counterexample. It is not necessary to understand the proof in order to know that it is correct.
At least in principle. Practically, any system as least as powerful as first-order logic is undecidable, so there can never be any computer program that would be able to do this perfectly. It might be that some day, an AI could be just as good as a trained mathematician, but so far, it doesn't look like it.
At least in principle. Practically, any system as least as powerful as first-order logic is undecidable, so there can never be any computer program that would be able to do this perfectly. It might be that some day, an AI could be just as good as a trained mathematician, but so far, it doesn't look like it.