Researchers have found that the hallucination rate of large language models when answering clinical questions can reach 5% to 15%.