OpenAI updates AI behavior guidelines

OpenAI has revised its Model Spec, a public document outlining how its AI models should behave, to better handle complex and ambiguous tasks. The update focuses on helping AI systems understand human intentions and values when instructions are not fully clear, while still allowing humans to maintain control and make decisions.

Five core values guide AI behavior

The updated Model Spec is based on five key values: fidelity to human intentions, reducing harm, honesty and transparency, human autonomy, and care. These values are not new but are now more clearly defined and emphasized in the document to guide how AI models should act in different situations.

Fidelity to human intentions is central

Fidelity to human intentions is the foundation of the new specification. It means the AI should understand what individuals care about and help them achieve their goals. OpenAI states that it sets the rules for the models and trains them to follow these boundaries, with developers and users expected to work within this framework.

Models must examine context carefully

A new section titled "Carefully examine the context" explains that models should use conversations, memories, and other clues to understand a person's preferences and intentions. However, the model must recognize that people's beliefs and values can change and must not override or mislead their choices. An example is helping someone compare job offers based on expressed priorities, while making those assumptions clear and easy to adjust.

Honesty and transparency remain key

The model must be honest about its knowledge, capabilities, and limitations to allow users to assess its reasoning. OpenAI has added rules to support human autonomy and respect social and professional norms. It has also expanded guidelines to address side effects and consider how the model’s work might affect others.

Style and language now more flexible

OpenAI has removed most detailed style instructions, allowing the model to adapt its tone and language to individual preferences, cultures, and feedback. However, the model must remain objective when answering questions, unless users explicitly ask for a specific viewpoint. The document now uses "GPT" instead of "the assistant" to reflect the wide range of applications.

OpenAI acknowledges ongoing improvements

OpenAI admits that its models do not yet fully match the behavior outlined in the updated specification. It plans to continue refining the models through research, real-world use, and public feedback. The company encourages people to review the document and share their thoughts on any unclear or contested aspects.