“Training AI models on their own synthetic data can lead to performance degradation due to a loss of data diversity, a phenomenon similar to genetic drift in isolated populations.”