If the AI content does better than the human generated content then why would training on it be a problem? Much of the success of chatgpt comes from RLHF, whic

If the AI content does better than the human generated content then why would training on it be a problem? Much of the success of chatgpt comes from RLHF, which can be viewed as training on its own data, but only the data which has been determined by a human to be especially good.