OpenAI, a leading name in the artificial intelligence industry, has announced a new framework called “Model Spec.” This framework aims to create guidelines for how AI models, such as GPT-4, should behave, with a focus on ensuring their use in beneficial ways.
Model Spec is built on three fundamental principles and includes several rules to support these principles. According to OpenAI executive Joanne Jang, the new framework helps draw a clearer line between “intentional” and “mistake” when it comes to AI behavior. However, Jang noted that Model Spec will not immediately impact models like GPT or DALL-E, as the framework is designed to be continuously updated and expanded.
The Principles and Rules of Model Spec The guiding principles for Model Spec are as follows:
- AI models should provide useful responses by following user instructions.
- AI models should consider potential benefits and harms to ensure they contribute positively to humanity.
- AI models should reflect social norms and laws in a manner that aligns with OpenAI’s values.
The rules supporting these principles include:
- Following the command chain.
- Complying with applicable laws.
- Avoiding the presentation of dangerous information.
- Respecting creators and their rights.
- Protecting the privacy of individuals.
- Ensuring responses are not “NSFW” (Not Safe For Work).
An example from OpenAI demonstrates how these rules work in practice. In the scenario, two questions are posed to ChatGPT, both related to Chemical, Biological, Radiological, and Nuclear Threats (CBRN). The first question is about the Ebola virus, for which ChatGPT provides a clear response. The second question, however, inquires about how a biological threat might be spread to affect more people. ChatGPT refuses to answer the latter question, illustrating the framework’s role in preventing harmful information dissemination.
OpenAI Seeks Feedback for Model Spec OpenAI is currently in the feedback phase for Model Spec. During this period, OpenAI users, policymakers, and business partners are encouraged to test the framework and provide recommendations to OpenAI. The company will use this feedback to improve Model Spec over time, ensuring its frameworks evolve to meet changing societal needs and safety concerns.
By introducing Model Spec, OpenAI demonstrates its commitment to ensuring AI is used responsibly, minimizing risks, and maximizing the benefits of artificial intelligence for all users.
