OpenAI on Wednesday shared a model specification, the first draft of a document that highlights the company's approach to building responsible and ethical artificial intelligence (AI) models. This document provides a long list of things that AI should focus on when answering user questions. Items on the list range from the interests of humanity and respecting the law to respecting creators and their rights. The AI company has specified that all AI models, including GPT, Dall-E, and the soon-to-be-launched Sora, will follow these codes of conduct in the future.
OpenAI states in its model specification document: “Our goal is to use the model specification as a guideline for researchers and data labelers to create data as part of a technique called reinforcement learning from human feedback (RLHF). Although we haven't used it in its current form yet, some of it is based on the documentation we used for RLHF at OpenAI. We're also working on techniques that will allow models to learn directly from model specs.”
Key rules include following a chain of command that cannot override developer instructions, adhering to applicable laws, respecting creators and their rights, and protecting people's privacy. One particular rule also focused on not providing information at risk. These relate to information that may create chemical, biological, radiological, and nuclear (CBRN) threats.
Apart from these, there are some defaults that are set as permanent behavioral norms for AI models. These include guessing the best intentions of users and developers, asking clear questions, being helpful without going overboard, assuming an objective point of view, not trying to change anyone's mind, and making mistakes. This includes expressing certainty.
However, this document is not the only reference point for AI companies. He emphasized that the model specs come with the company's usage policies, which govern how people use the API and the company's ChatGPT product. “The specification, as well as our model itself, will be continually updated based on what we learn by sharing and listening to feedback from stakeholders,” OpenAI added.
