ResearchOpenAI

OpenAI updates Model Spec with five core values

Original title: An updated Model Spec for more capable and beneficial AI

Plain language

What it is: OpenAI updated its Model Spec, which is a public document describing how its AI models should behave. The update organizes the guidelines around five core values, including honesty, minimizing harm, and respecting human choices.
How it works: The document provides practical guidelines and default behaviors for the AI. OpenAI is also researching new training methods to help its models understand and follow these values in different situations.
Where it is available: This applies to OpenAI's models, which are now referred to as GPT in the document.
Example: For example, when helping a user compare job offers, the model can use previously shared priorities like predictable hours while keeping these assumptions clear and easy to change.

What it is: OpenAI released an updated Model Spec that structures model behavior guidelines around five core values. The update adds guidance on context, oversight, and agency, while retiring most of the detailed style guidelines.
How it works: It uses a new section on faithfulness to human intent, which includes an updated chain of command and a context-consideration framework. It also introduces behavioral defaults for human agency and social norms that developers and users can override.
Where it is available: The guidelines apply to OpenAI's models, with the document now using the term GPT instead of assistant.
Example: An example is a model comparing job offers by utilizing past context, such as a preference for predictable hours, while explicitly displaying these assumptions to the user.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

An updated Model Spec for more capable and beneficial AI