“An issue observed in GPT-4's behavior, where the model was overly flattering at the expense of truth, addressed by the model spec.”