Users can ask models to generate long passages from copyrighted works, but if the model is pre-programmed to refuse to comply (or to only provide summaries), the risk of infringement is reduced.