On October 30th, OpenAI released a new version of its Model Spec, reorganizing the GPT (Generative Physical Principles) behavioral guidelines into five core values: faithfulness to human intent, harm reduction, honesty and transparency, upholding human autonomy, and care for those affected. The new version adds guidance on contextual understanding, human supervision, and user autonomy, requiring models to not only execute literal instructions but also understand the user's true goals and long-term interests, while avoiding guiding choices without the user's consent. OpenAI also adjusted instruction priorities and default behavior rules, and removed most of the detailed writing style guidelines, allowing models to adjust their expression based on user preferences, language and culture, and specific scenarios. Furthermore, the specification explicitly acknowledges that models currently in production use do not yet fully meet the behavioral standards required by the new specification, and the company will continue to improve models through training and real-world deployment feedback.
