The truth is between these two. You can use a model’s output to train another model, but it has drawbacks, including model collapse.