I don't think it says that? The paper argues very specifically that the RL post-training is structured to produce hallucinations and suggests ways to change it to minimize hallucinations