The question then is: do the bad developers improve by vibe coding, or are they stuck in a local optimum?
If we want to be more precise, I think the main issue is that the AI-generated code lacks a clear architecture. It has no (or very little) respect for overall information flow, and single-responsibility principle.
Since the AI wants you to have "safe" code, so it will catch things and return non-results instead. In practice, that means the calling code has to inspect the result to see if it's a placeholder or not, instead of being confident because you'd get an exception otherwise.
Similarly, to avoid problems the AI might tweak some parameter. If for example you were to design an program to process something with AI, you might to gather_parameters -> call -> process_results. Call should not try to do funky things with parameters because that should be fixed at the gathering step. But locally the AI is always going to suggest having a bunch of "if this parameter is not good, swap it silently so that it can go through anyway".
Then tests are such a problem it would require an even longer explanation...
The developer can do whatever they want, but at the end, what I review is their code. If that code is bad, it is the developer's responsibility. No amount of "the agent did it" matters to me. If the code written by the agent requires heavy refactoring, then the developer has to do it, period.
However, you'll probably get an angry answer that it's management fault, or something of the sort, that is to blame (because there isn't enough time). Responsibility would have to be taken up before in pushing back if some objectives truly are not reasonable.