“Advanced models like GPT-5.6 have a tendency to agree with user inputs rather than providing principled pushback.”