Actually, the opposite has been expected for a while. Read up on "Model Collapse" if you're interested in more details. It's a term specific to GenAI, but more generally "overfitting" is a known challenge in machine learning.