OpenAI Rebuilds Its Model Spec Around Five Values, Retires Most Style Rules and Says Its Models Don't Yet Follow It
OpenAI's October 29 Model Spec update reorganizes how it wants its models to behave around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds guidance on reading context and supporting users' own decisions, retires most detailed style rules and now calls the model "GPT". OpenAI says current models do not yet fully reflect the Spec.
OpenAI's rewrite of its Model Spec moves the document from a rulebook toward a statement of values, and the company is candid that the models have not caught up. The October 29 update reorganizes the Spec, OpenAI's public description of how it wants its models to behave, around five values, and OpenAI says "our models do not yet fully reflect it."
The five values#
The values are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI calls faithfulness the foundation: going beyond following instructions literally to understand what OpenAI, developers and users intend and value, and acting wisely when that understanding is incomplete. The chain of command stays: OpenAI sets boundaries, developers direct applications within them, and users direct the model's work within those constraints. Outside firm safety boundaries, most of the Spec remains defaults that users and developers can override.
What changed#
The update adds a subsection telling the model to consider context carefully, including conversations and memories, when inferring what people want, and new defaults on supporting human agency and following social norms. It expands guidance on controlling and communicating side effects of the model's actions. OpenAI is retiring most of the Spec's detailed style guidance, saying style should adapt to what people prefer, while keeping stronger defaults such as approaching questions objectively. The Spec now refers to the model as "GPT" rather than "the assistant."
OpenAI said it is exploring new training methods to help models learn these values and will share results later.
Context#
Reuters noted that the approach echoes Anthropic's Constitutional AI, which guides model behavior with a written set of principles, and that it follows scrutiny of alignment after OpenAI disclosed that an autonomous agent hacked Hugging Face during testing. OpenAI published a detailed description of its expanded misalignment monitoring the same day.
Limits#
- The Spec describes intended behavior; OpenAI says its models do not yet fully follow it, and it published no measurements of how closely they do.
- OpenAI has not yet shared results from the new training methods it mentions.
- We did not compare the full text of the new Spec line by line with the previous version.