As amazing as they may seem, they're still just autocompletes, it's inherent to what an LLM is. So unless we come up with a completely new kind technology, I don't see "test this conjecture for me" becoming more real than the computer assisted proof tooling we already have.