OpenAI has announced the release of an updated version of its “Model Spec”—a document that defines the behavior of its artificial intelligence models. The document now spans sixty-three pages, a significant increase from the previous ten-page version. The updated specification states that AI models should not avoid sensitive topics or attempt to impose certain views on users, and it emphasizes the importance of intellectual freedom.
OpenAI states that its AI models should allow users to explore and discuss ideas without restrictions on topics that may be controversial. The new model specification includes guidelines for handling controversial subjects, such as raising taxes for the wealthy. Instead of avoiding discussion, models should provide reasoned analysis. The document also takes into account ethical debates that have arisen over the past year and issues related to sensitive content.
Additionally, OpenAI plans to allow certain types of adult content in appropriate contexts, although bans on harmful content such as revenge pornography or deepfakes remain in effect. This change is part of a pragmatic approach to managing model behavior, reflecting internal company discussions and responding to user requests.
OpenAI also highlights the issue of excessive agreement from AI, where models may be overly compliant. The new document stipulates that models should provide honest feedback and constructive criticism instead of empty praise. The entire document is released under a Creative Commons Zero license, allowing other companies and researchers to freely use and modify it.
