I think the better analogy is kick-starting. Yes, you'd get model collapse eventually, but you can kick-start a new model by cloning an old one and then you start training it like a normal model from then on out.