Did it not turn out that AIs creating output based on AI-created output, iteratively, pretty quickly turns into nonsense? That would suggest that we'll need humans always to provide training material, and for that training material to be meaningful, it'll have to be created by humans who know what they're doing.