You still need human writers to train the model. You can't train an AI on AI output, for the same reason you can't keep making a photocopy of a photocopy.

The first generation of AIs were able to train on freely available text on the Internet. They also didn't have to worry about AI output polluting their input data. That isn't the case anymore. If you use random internet posts to train your AI, you're including a lot of AI.

AI is good for non-critical tasks. It still suffers from the "hallucination" problem, where it gives a completely wrong answer with supreme confidence.

You should read about how AI corporations are now hiring an army of writers to help train their AIs.