OpenAI's Approach to the Model Spec

As AI systems become more advanced, a clear public framework for how models should behave is needed. Through the Model Spec, OpenAI's official framework defining model behavior, the company specifies how models should follow instructions, resolve conflicts, respect user freedom, and act safely.

The Model Spec is not merely an internal training tool but is intended to be a public document that users, researchers, and policymakers can read, review, and discuss. Rather than claiming that current models are perfect, it serves as a benchmark setting goals for where future models should head.

This framework is a core element of OpenAI's strategy for safety and building responsible AI.

• Preparedness Framework: Managing risks and establishing safeguards for frontier models
• Model Spec: Defining how models should behave in various situations
• AI Resilience: Addressing the broader challenge of reducing societal disruption and maximizing the benefits of AI

Rather than being a fixed document, the Model Spec follows an Iterative Deployment approach, evolving through actual deployment and feedback. The three core goals for system-level optimization are as follows.

• Iterative deployment of models that support developers and users
• Preventing actions that cause serious harm to users or others
• Maintaining OpenAI's operating license