Thinking the user of AI can just spec a goal and then the AI automagically produces perfect results is just the pipe dream of the lazy who won't ever have a clue about how to verify the resulting system, much less the system that produced it. Good luck with that!
Software is about 100% correctness, which is damn-near impossible. And when it fails, it will fail at the worst time and in the worst way possible. It's just a matter of when, not if. We can't even remotely do that (for nontrivial systems), but we're gonna make a machine that can do it?!
Shhheeeeeiiiitttttt.