A group of Stanford researchers found that large language models can propagate false race-based medical information.