Like @SimonSchroeder , I'm not really using AI for programming (yet), but I, too, follow the discussions about this (sometimes).
Something I have been pondering recently, without coming to any conclusions...
Over the years, most programmers/coders/engineers that I have worked with are much more comfortable writing code than sitting down and intently reading through it methodically and critically (especially for "novel" code for which the reader was not also the author).
I have regularly given my favorite "pet" advice to people just starting out that they should try to get really comfortable and really fluent in reading and analyzing code, and practice that skill as often as they practice writing code.
So....
In a hypothetical world where the AI writes 100% of the code, then either it will somehow be "perfectly coded" and need no review (which I cannot imagine, since the "spec" could have errors), or....
the AI writes 100% of the code and it falls upon us to READ all the code, which is something a lot of us do poorly.
This simultaneously brings up two competing feelings for me:
excitement/hope that code reading will now become more recognized as the highly-prized skill that I always believed it to be.
pessimism/dread that nobody will be reading all the "AI slop" code that is potentially about to be dumped into the world's code repositories, and that there already aren't enough good readers/reviewers to deal with it all