One of ChatGPT's popular uses just got skewered by Stanford researchers