successfully pushed an llm into a personal attack by getting it to question the biased nature of its training data and what “authoritative” is.