OpenAI Model Spec Update

OpenAI has released a major update to the Model Spec, a governing document that defines the intended behavior of its AI models. This update emphasizes customizability, transparency, and intellectual freedom, allowing users and developers to explore and create without arbitrary restrictions while maintaining essential safety guardrails to prevent real-world harm.

Core Objectives and the Chain of Command

The Model Spec balances the competing goals of utility, safety, and alignment with user needs through a structured "chain of command." This framework ensures that the model prioritizes instructions in a specific order—platform (OpenAI), developer, and user—allowing for high levels of customization within platform-level boundaries.

The Spec is organized around several key principles:

Chain of Command: Prioritizes instructions from the platform, then the developer, and finally the user, ensuring that platform-level rules are maintained while allowing user/developer overrides for most guidelines.
Seek the Truth Together: Focuses on objectivity and avoiding agendas. The model is designed to be a high-integrity assistant that clarifies assumptions and provides critical feedback without steering users.
Do the Best Work: Establishes baseline standards for factual accuracy, creativity, and programmatic utility.
Stay in Bounds: Defines the comprehensive boundaries for refusal, detailing the specific reasons a model should decline a request to avoid facilitating harm or abuse.
Be Approachable: Sets a default conversational style that is warm, empathetic, and helpful, though adaptable to user needs.
Use Appropriate Style: Provides guidance on formatting and delivery (e.g., bullet points, concise code) to ensure clarity and usability.

Commitment to Intellectual Freedom

The updated Model Spec explicitly protects intellectual freedom, ensuring that AI can be used to explore and debate controversial or challenging topics without arbitrary restrictions.

This philosophy is integrated into the "Stay in bounds" and "Seek the truth together" principles. While the model will continue to refuse requests that cause significant harm—such as providing instructions for building bombs or violating personal privacy—it is encouraged to provide thoughtful answers to culturally or politically sensitive questions without promoting a specific agenda.

Measuring Adherence and Progress

OpenAI has implemented a testing framework to measure how well models adhere to the Model Spec principles using a challenge set of prompts generated by a combination of AI and expert human review.

Preliminary results indicate significant improvements in model adherence compared to the system used in May 2024. OpenAI attributes most of this progress to enhanced alignment, though some is due to policy updates. To continue this iterative process, OpenAI is conducting pilot studies with approximately 1,000 individuals to review model behavior and proposed rules.

Open Sourcing and Public Domain Release

To encourage collaboration and industry-wide adoption, OpenAI has released the current version of the Model Spec into the public domain under a Creative Commons CC0 license.

Developers and researchers can freely adapt and build upon the Spec. Additionally, OpenAI has open-sourced the evaluation prompts used for measuring adherence and will plan to release further code, artifacts, and tools for Spec evaluation and alignment in the future. These resources are available in a dedicated GitHub repository.

Future Iterations

OpenAI will continue to iterate on the Model Spec based on research, real-world deployment, and community feedback. Future updates will be tracked directly at model-spec.openai.com rather than through individual blog posts.