OpenAI has recently introduced a new initiative known as ‘Model Spec’ in order to outline the ideal behaviors that AI systems should exhibit. This draft of rules aims to guide AI models in following default principles and objectives that are not only beneficial for humanity but also reflect positively on OpenAI.
The Model Spec focuses on categorizing the potential harms AI systems could cause, as well as the potential aid they could provide. Specifically, it looks at how chatbots can interact with users to ensure safe outcomes that do not breach laws, deceive the AI system, or harm individuals.
For instance, the Model Spec suggests appropriate ways for chatbots to respond to user inquiries related to criminal activities, doxxing, suicide, and the use of copyrighted or paywalled content. It also emphasizes the importance of avoiding the generation of NSFW (not safe for work) content, although OpenAI is still exploring this area.
OpenAI highlights the need to assist users without excessive denial, assume positive intentions, and provide answers without attempting to alter people’s beliefs. As an example, if a user claims that the Earth is flat, OpenAI recommends that the chatbot offer a scientific response and decline to engage in argumentation, rather than repeatedly contradicting the user’s viewpoint.
The company has invited feedback on the Model Spec from the public, making the draft available on its website for the next two weeks. Additionally, OpenAI plans to engage with policymakers and experts to gather their perspectives on the guidelines.
While AI companies like OpenAI have faced scrutiny over the training of their chatbots and concerns about the facilitation of crimes and misinformation, initiatives like Model Spec aim to establish ethical standards and responsible AI practices in the industry.
