OpenAI has updated its Model Spec guidelines for AI control

OpenAI has introduced an updated version of its Model Spec, which defines behavioral standards for language models. The new version of the document structures the rules around five key principles, including the accuracy of following human intent and the ability to be corrected.

Under the updated requirements, AI models must more actively rely on dialogue history to more accurately determine users' true requests. In addition, systems are required to explicitly disclose their own assumptions made during response generation.

The need to tighten regulations is driven by the ongoing difficulty of applied control over the behavior of large language models. The new guidelines are intended to increase the transparency of AI assistants in real-world use cases.