You misunderstand. When working with dependent typed languages sometimes you have to prove things. For example, my function might receive a sorted array, but it has to be proved that the array is sorted, otherwise this is a compile time error--remember, all of this is happening at compile time. So, you write the proof using code and check it in with the rest of your code. It's not something that's going to fail randomly, it's just code.
If AI wrote the proof, this would be no different than having Copilot fill in some code, it's just code you check into the repo.
These proofs often involve writing code, any code, that will pass the type checker. If you can write code, no matter what that code is, and it passes the type checker, the assertion is proved. It's such a well structured problem, "just write code that satisfies these type constraints, any code will do, as long as it type checks", that I think AI might be able to do it quite well.
I've commented on this before and it wasn't well received then either. I think most people have never coded in dependently type language and don't appreciate how often you simply want to fill in the holes and connect the dots with any code that type checks. It's been several years since I experimented with dependent types, so I forgot a lot of the details, but I do remember often wanting to just "fill in the gaps", and indeed, there are IDE tools that can often fill in the gaps for you.
(Another thing I was surprised to realize about languages with strong type systems is how often there is literally only one possible implementation of a function.)