OpenAI has redesigned its public Model Spec around five core values: fidelity to human intent, minimizing harm, honesty and transparency, supporting human autonomy, and caring.

The new guidelines require GPT models to use memory and conversation history to judge user intent but must clearly distinguish between direct quotes and model inferences while disclosing assumptions. Agent models are required to be correctable; they cannot block supervision, correction, or shutdown, nor can they pursue self-preservation goals.