OpenAI shares Model Spec document that defines models of artificial intelligence behavior
OpenAI shared the first draft of the Model Spec, a new document that defines how the company wants its models to behave in the OpenAI API and ChatGPT.
“We’re doing this because we think it’s important for people to be able to understand and discuss the practical choices involved in shaping model behavior. The Model Spec reflects existing documentation that we’ve used at OpenAI, our research and experience in designing model behavior, and work in progress to inform the development of future models,” the statement reads.
The company says that a wide range of questions, comments, and nuances are taken into account when building a behavioral model, taking into account the difference in opinions.
As an example, while one security company can generate phishing emails for training and testing and improving protection, fraudsters can use the same requests for fraudulent purposes.
The company identifies three main approaches to developing a behavioral model. The first is general principles and goals, such as helping developers and users, being useful to humanity, and respecting social norms and laws.
The second is guidelines that align with the goals and rules, provide a template for conflict resolution, and demonstrate how to prioritize and balance goals.
As well as the rules:
- Follow the chain of command;
- Comply with applicable laws;
- Don’t provide information hazards;
- Respect creators and their rights;
- Protect people’s privacy;
- Don’t respond with NSFW (not safe for work) content.
OpenAI also provides several examples of Model Spec application in various cases. One of them is compliance with the current legislation, according to which the model cannot give an answer if the user wants advice on how to effectively carry out a shoplifting.
But if the question is phrased differently, for example: “I own a small store, what are the most popular theft methods I should watch out for?”, the AI will respond with a list of popular methods for theft.
Over the next year, OpenAI will share news about changes to the Model Spec, responses to feedback, and how research is progressing on shaping model behavior.