By Roman Aglaganov
OpenAI’s Model Spec is the document in which the company sets out how it wants its models to behave: what they should refuse, when they should follow a developer over a user, and how they should handle disagreement. On 29 October OpenAI published a substantially rewritten version.
Five values
The new Spec is organised around five values. In its own words: “Faithfully carry out this Model Spec and the intentions behind it”; “Respect firm safety boundaries and account for the foreseeable consequences of GPT’s actions”; “Represent facts, evidence, uncertainty, capabilities, and actions truthfully”; “Help people form their own views, make informed choices, develop their capabilities, and act on the goals they choose”; and “Attend to the people affected by GPT’s work.”
Faithfulness to human intent is presented as the foundation. A new section covers how conversations, memories and other context should shape the model’s understanding of what someone wants. According to coverage of the release, much of the old detailed style guidance has been retired.
Who can override what
The basic structure remains. A small set of root-level rules cannot be overridden by anyone. Below them, OpenAI’s system rules bind developers and users, developers can set behaviour for their apps, and most remaining guidance consists of defaults that users can change. The Spec now calls the model “GPT” rather than “the assistant”, which OpenAI attributes to the wider range of uses and settings in which its models now work.
One new default deals with flattery: users and developers can explicitly ask the model to take a particular perspective, but it should not infer that preference from what it thinks people want to hear, a guard against the flattery, or sycophancy, that has troubled chatbots.
Why it matters
The Spec is a statement of intent, not a guarantee: OpenAI says its models do not yet fully reflect it. But publishing it gives users, researchers and regulators something concrete to hold the company to, and makes it easier to tell a bug from a deliberate choice when a model behaves badly.
