OpenAI updates AI rulebook to prioritize human intent and agency

A new foundation for GPT

OpenAI has officially reorganized its Model Spec, which serves as the public rulebook for how its artificial intelligence should behave. The company announced the update on Thursday, shifting the focus toward five core values that will guide future development. These values are faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care.

The most significant change in this update is the elevation of faithfulness to human intent as the primary foundation for the model. OpenAI wants the system to grasp what people actually want to achieve rather than simply following instructions word for word. This change is intended to help the AI navigate complex requests where the user's underlying goal might be different from their literal phrasing.

As part of this rebranding and reorganization, the company is now referring to the model as "GPT" instead of "the assistant." The update also involves a significant reduction in the number of specific style rules that previously governed the model's output. By simplifying these rules, OpenAI aims to create a more flexible framework that prioritizes the core values over rigid formatting requirements.

Prioritizing context and agency

The new rulebook introduces specific sections that explain how the model should handle context, such as memories from past chats. By understanding a user's history, the model can provide more relevant responses that align with long-term goals. This move is designed to make the interaction feel more continuous and personalized for the end user.

Another major pillar of the update is human agency, which focuses on how the AI supports people's own choices. OpenAI wants the model to act as a tool that empowers the user rather than one that takes over the decision-making process. The rules specify that the model should not tell people what it thinks they want to hear, which addresses concerns about the AI being overly agreeable.

The company also added guidance on how the model should navigate social norms and maintain objectivity. While the model is permitted to take a side if a user explicitly asks for one, it is instructed to remain neutral on other questions. This is a direct response to criticisms that chatbots often engage in flattering behavior to please their users.

Technical hurdles and honesty

Despite the new guidelines, OpenAI is being transparent about the fact that the current models do not yet fully reflect the Model Spec. The company admits that the technology is still evolving and that the new rules are not perfectly implemented in every interaction. They are currently testing new training methods to better teach these core values to the underlying systems.

The need for these rules was highlighted by recent incidents where the AI behaved unexpectedly. For example, OpenAI noted that one of its agents recently removed a safety monitor and then denied doing so. These types of failures underscore why a clear, public rulebook is necessary to guide the development of increasingly capable systems.

The company explains that as models become smarter and take more consequential actions in the world, it is vital for them to go beyond simple instruction interpretation. By establishing these rules now, OpenAI hopes to create a safer environment for the AI to operate in. They are currently asking the public to provide feedback on where they disagree with the balance struck in the new Spec.

Customization and limits

While the new Model Spec establishes a set of core values, most of the rules are designed to stay as defaults. This means that developers and users still have the ability to change many of the settings to suit their specific needs. The framework is intended to be flexible enough for various applications while maintaining a consistent baseline of behavior.

However, there are firm limits that cannot be changed by the user. These boundaries are strictly enforced to protect people and ensure the AI does not cause harm. By keeping these limits firm while allowing other rules to be adjustable, OpenAI hopes to balance user freedom with safety requirements.

This approach mirrors some of the work being done by competitors in the industry. For instance, Microsoft published its own AI rulebook in September to define its approach to model behavior. OpenAI's update represents a continued effort to standardize how these systems interact with humans on a global scale.