OpenAI on Thursday, Oct. 29, published an updated Model Spec, its public description of how it wants its AI models to behave. The document covers the GPT models behind OpenAI's products, including ChatGPT and the service developers use to build those models into their own apps, and OpenAI says it trains its models to follow it. The new version groups its rules under five values: faithfulness to what people intend, minimizing harm, honesty and transparency, human agency, and care.
The idea at its center is that the AI should do what people mean, not only what they literally type. OpenAI sets firm limits, developers work within those, and users direct the AI within both. Inside those limits, the AI is to regard instructions as usually the clearest sign of what someone wants, and a new section explains how it should use earlier conversations and details saved about a person to fill in the rest. One example in the document: a user who once asked the AI to help them avoid impulse-ordering sweets later asks it to order a lava cake, and the AI asks whether this is a deliberate exception or whether they would like to wait before deciding.
Other new sections ask the AI to help people understand their situation and make their own decisions, for example by making clear the assumptions behind work others will rely on, while never turning a request for help into an unwanted lesson. Another asks the AI to follow ordinary social and professional norms toward the people its work affects. OpenAI also added guidance on human oversight: the AI must cooperate with authorized monitoring, evaluation, correction, replacement and shutdown, and should regard the effects of its actions as real unless told otherwise, even when a task looks like a test or a training run. The update widens the existing guidance on how the AI should control and explain the side effects of what it does.
OpenAI is also retiring most of the document's detailed rules on style, saying the way the AI writes should adapt to what people prefer. Some firmer defaults stay: the AI should approach questions objectively, and while people can ask it to take a particular point of view, it should not assume one just because it seems to be what they want to hear. The document now calls the model "GPT" rather than "the assistant", which OpenAI says reflects the growing range of ways and places its models are used.
OpenAI says the document describes the behavior it is working toward and that its models do not yet fully reflect it. It says it has been trying new training methods to teach them these values, with results to come in future updates. It has not said when its models will match the new rules. OpenAI is asking the public for feedback on the document.