I've maintained for years that my engineering job is, sarcastically, "to move entropy around." This is no different, as we've squeezed the envelope at one end only to balloon it at the other.

I would add that once you're all-in on codegen for a project that escapes human comprehension, you have to keep using more codegen to maintain it. It effectively hands off the whole game to the robots, and probably keeps growing in scale all the time. By way of your observation of increasing cost of human involvement, it resembles a massive tech-debt scheme to me. Which is to say that it'll work impressively well in the short term until someone has to get their hands dirty with the gordian knot of code that has been quietly cooked by LLMs for years. At that point, the bill comes due.

I think this is the best argument for human-in-the-loop, or even my personal favorite "LLM as an advisor or intern" instead. I've also had traction with AI codegen for disposable/single-use programs, where I just need to know if something works.