“Since its early models, OpenAI has intentionally avoided supervising a model's thinking process to prevent it from learning to generate appealing but dishonest reasoning.”