Introducing the Model Spec

The Model Spec is a document that defines how models should behave in OpenAI's API and ChatGPT. It specifies behaviors that critically affect user interaction—such as the model's tone, personality, and response length—in order to increase transparency in the model development process and spark social discussion.

Through a recent update, OpenAI reaffirmed its policy of reducing arbitrary restrictions while strengthening customizability, transparency, and intellectual freedom, all while maintaining guardrails to prevent real-world harm.

The guidelines are broadly composed of three layers.

• Objectives: Broad principles to help developers and users, benefit humanity, and reflect OpenAI's values.
• Rules: Guidelines for safety and legal compliance, including complying with laws, preventing information hazards, respecting copyright, protecting privacy, and restricting NSFW content.
• Default behaviors: Templates for determining priorities in conflicting situations, such as assuming positive intent from the user, asking questions when needed, and maintaining an objective viewpoint.

OpenAI plans to use this document as a guideline for researchers and trainers performing RLHF (Reinforcement Learning from Human Feedback). It is also exploring the possibility that future models could learn directly from this Model Spec.