The biggest problem in my head with AI generated code is that its mistakes are subtle but can still be critical. There will be a point where people don't understand generated code and just leave it unmodified allowing other code to pile up and depend on it. At that point you no longer have a bug, but a new feature. Also, AI doesn't grasp things on a big scale, just shits out output with highest score. This doesn't mean output is a great fit for your project or for upcoming plans.