it is better able to generate or augment its own training data

And here starts the problem. The "good" training data is used, once there's an imbalance of human training data against spewed out GPT plausibility generated data it's a race to the bottom in data quality but also model quality.