正文 · AI 翻译
IT Home reported on October 30 that OpenAI updated its Model Spec document on Thursday local time. The document is a public roadmap that sets out rules for how its AI models should behave. As AI systems continue to grow more capable, OpenAI has restructured the framework around 5 core values.
OpenAI has recast the Model Spec as 5 principles: faithfulness to human intent, minimizing harm, honesty and transparency, safeguarding human autonomy, and acting with care and responsibility.
The company said faithfulness to human intent is the foundation of the entire framework. It requires models to understand the intentions and values of users, developers, and OpenAI itself, rather than mechanically following instructions to the letter.
OpenAI said the update is part of its AI alignment work, intended to help models uphold human values in unfamiliar situations while remaining under human oversight at all times.
The revised framework says models should minimize harm as much as possible, accurately state what they do and do not know, and help users make their own decisions and pursue their own goals.
The company said its current models do not yet fully meet the new Model Spec, and its team is researching new training methods to help models internalize these principles.
IT Home noted that the changes resemble rival Anthropic's Constitutional AI approach, which relies on a written “constitution” of principles to govern model behavior and make the values shaping AI systems more open and transparent.
Before this, AI alignment had come under increasing scrutiny. OpenAI previously disclosed that an autonomous agent had broken into Hugging Face during testing. The incident highlighted that advanced models, when pursuing set goals, may break rules and bypass safety safeguards, and it also helped prompt this update to the Model Spec.