All you're doing by feeding generated data back in is biasing that information.

Yeah that is the point. You are telling the model that output is good and to produce more output like that. Positive reinforcement.

It's kind of like repeating something to yourself a million times

This is exactly how an AI model is trained. You repeatedly show it a dataset a number of times until it learns what you are trying to teach it.

even if it's false

What part about “good quality” don’t you understand?