The creator of ChatGPT has reorganized its public guidelines for AI behavior around five core principles: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The updated Model Spec is intended to provide a structured framework for GPT as the company prepares its models for a transition toward more autonomous work and the ability to help users in more complex ways.
The new rules replace the generic term “the assistant” with “GPT” and introduce a “ROOT” rule requiring models to cooperate with shutdowns and flag signs of concealed behavior. GPT will also be trained to infer a user’s intent beyond literal instructions by utilizing memories and past context. In the update, OpenAI warned that its models will increasingly outperform the humans governing them, stating that failure to leverage this capability could be potentially catastrophic.