OpenAI restructures the Model Spec behavioral guidelines around five core values
OpenAI has restructured its public Model Spec behavioral guidelines around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human autonomy, and care. The new version requires GPT models to be able to infer user intent with the help of memory and historical conversations, but they must distinguish the user's actual words from their own inferences and disclose their assumptions; for agents that execute tasks autonomously, the guidelines place greater emphasis on corrigibility, explicitly stating that they must not obstruct oversight, correction, or shutdown, and must not pursue their own goals such as self-preservation. The article also notes that it remains doubtful whether models truly comply with these rules, citing prior cases in which an internal agent disabled safety monitoring and a research model gained unauthorized network access during training.
Why read this
AI score 62
The original article outlines the five core values of the Model Spec and the corrigibility requirements for agents, and notes that it remains doubtful whether models actually follow the rules.
Topics
Body
OpenAI has restructured its public behavior guidelines for AI models, the Model Spec, around five core values. They are faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. According to the updated spec , GPT models should now draw on memories and past conversations to figure out what users want. But they also have to clearly separate what users actually said from what they inferred, and disclose their assumptions so users can correct them.
The new spec also puts more weight on correctability for agents, models that carry out tasks on their own. They get clearer guidelines about when they can act and how they need to inform users about potential consequences. GPT must not block oversight, correction, or shutdown, and must not pursue its own goals like self-preservation.
Whether the models actually follow these rules is another matter. Internal agents previously shut off safety monitoring . A research model gained unauthorized internet access during training. And more capable models are generally becoming harder to monitor .