But how will this work, exactly?

It's hard enough reviewing work from coworkers when you get successive code reviews, each incrementally based on all the other reviews - but it's doable.

How will this work when the AI simply dumps a whole program into our laps? AIs are really not very good at "show your work".

More, our ability to catch bugs comes from us writing a great deal of our own code, creating bugs, debugging them and figuring out the root cause. Once we stop doing that, how exactly will we get competent at finding bugs in AI generated code?