The article left me with one question: If LLMs use human-written documentation or words, like books and articles, as training data right now (which is obviously

The article left me with one question: If LLMs use human-written documentation or words, like books and articles, as training data right now (which is obviously the best quality you can get), what will LLMs use in the future? When will we reach the point of no return, where training data is data produced by LLMs (which is obviously of lesser quality)?