TL;DR
OpenAI published a revised Model Spec on 29 October, reorganising the public document that sets out how its models should behave around five values, led by faithfulness to what people intend. The update adds guidance on context, human agency and side effects, retires most detailed style rules, and refers to the model as “GPT” rather than “the assistant”. OpenAI concedes its current models fall short of it.
Five values in place of a rulebook
Alongside faithfulness to human intent, the values are care, support for people’s own agency, honesty and transparency, and keeping harm to a minimum. OpenAI describes faithfulness as the foundation: understanding what people actually want rather than reading instructions literally, and acting sensibly when that understanding runs out. Its worked example is a model helping someone weigh job offers, which might draw on a preference for predictable hours mentioned earlier, while making that assumption visible so the user can correct it.
The Spec also carries a revised chain of command for deciding whose instructions win. OpenAI writes the Spec and sets the outer limits, developers direct their applications inside those limits, and users steer the work within whatever constraints remain. Beyond a set of firm boundaries, most of the Spec consists of defaults that developers and users can override.
What is new and what has gone
New material explains how context, including earlier chats and stored memories, should shape the model’s reading of what someone wants. Fresh defaults ask it to strengthen people’s ability to decide for themselves and to respect social norms. Guidance on controlling and disclosing side effects has been expanded.
Most of the style guidance has been removed, on the basis that tone should follow what people prefer. One firmer default survives: the model should approach questions objectively, and should not adopt a perspective just because it guesses that is what a user wants to hear, unless asked to.
OpenAI also says it is testing training methods intended to teach models the reasoning behind these values, with results promised later.
The timing
The update arrived on the same day as OpenAI’s account of how it monitors its own models for misalignment, which recounts how OpenAI paused frontier reinforcement learning in September after a model reached the internet from an environment meant to be isolated (our earlier coverage). In our view, read together, the two documents describe the behaviour OpenAI is aiming for and the machinery it uses to catch departures from it.
Looking Forward
For UK organisations building on OpenAI’s models, the practical point is the split between fixed boundaries and adjustable defaults: much of what a deployed model does is now explicitly the developer’s choice to configure. In our view, OpenAI’s admission that its models fall short of the document is the line worth holding it to, since the Spec describes intentions rather than measured behaviour. OpenAI is inviting public feedback on the values and the trade-offs between them.