They already do. All the current leading edge models are heavily trained on synthetic data. It's called textbook learning.