How will they be able to review AI generated code if they don't understand anything beyond the basics?
The same way most of us review our compiler generated code today (ie not at all). If it works it works, if doesn't we fix the higher level input and try again. I won't be surprised if in a few more generation the AI will skip the human readable code step and generate ASTs directly.
How can I visit this fantasy world of yours where LLMs are as reliable and deterministic as compilers and any mistakes can be blamed solely on the user?
Wait 20 years.
It's really easy to make unsubstantiated claims about what will happen decades from now, knowing your claims will be long forgotten when that time finally comes around.
We for example have a whole generation of programmers who have no idea what the difference between a stack and a heap is and know nothing about how memory is allocated. They just assume that creating arbitrarily complex objects and data structures always works and memory never runs out. And they have successful careers, earning good money, delivering useful software. I see no reason why this won't continue.