Researchers tested five popular large language models to find they all produced answers that lacked nuance or were sometimes outright wrong.