Your assumptions are wrong. AI models do not have equal generation and discrimination abilities. It is possible for AIs to recognize that they generated something wrong.

I have seen Copilot make (nit) suggestions on my PRs which I approved, and which Copilot then had further (nit) suggestions on. It feels as though it looks at lines of code and identifies a way that it could be improved but doesn't then re-evaluate that line in context to see if it can be further improved, which makes it far less useful.