Training neural network is using compute to find common patterns in input data. Its process of compressing what is common between each provided input segment into a generalized form. This generalized form can be then used to generate new data.
As opposed to popular view, where training AI on AI generated data is seen as providing worse output, the reality is opposite:
When we create generalization of some dataset, we reduce noise and extract what is important. When we train new AI on AI generated data, we generalize generalization, which results into a better generalization of original data.