OpenAI Shares Secret Rules Behind ChatGPT Behavior

OpenAI has shared a draft of some of the guidelines and rules structuring its generative AI models behind its API and ChatGPT. This Model Spec outlines the direction OpenAI is pursuing as it tries to standardize and improve the behavior of its models while also soliciting public comment and discussion on the ethical and practical implications of generative AI interactions.

Model Spec

The Model Spec serves as a kind of blueprint for how OpenAI wants its models to behave and the operational principles for any large language model (LLM) developed by OpenAI. The main facets of the Model Spec are objectives, rules, and default behaviors. Objectives provide a broad framework, aiming to assist users and developers while adhering to social norms and laws. The rules specify what that looks like in practice with regard to legal compliance, intellectual property rights, privacy, and ensuring a response from the API or ChatGPT is appropriate. The default behaviors expand on those rules for when the situation is more complex, with practical guidelines on how AI models should manage conflicts, maintain neutrality, and interact with users.

The Model Spec has a notable range of what these principles look like. They include assessing developer intent to decide how direct an answer is and whether the response to a question about how to contact someone would violate anyone’s privacy. It also tackles how to deal with users indenting to manipulate the model. ChatGPT or a bot built with OpenAI’s API might just not respond to questions that are deliberately provocative or involve sensitive topics, or kick the user to a human agent to handle.

The document consolidates existing practices while pulling from research input from experts in AI, ethics, and related fields. While any generative AI model would likely have some variation of a Model Spec, it is unusual to see it shared publicly, especially when the models themselves are not open-source. However, OpenAI ties the release to its larger strategy and philosophy as a company and is inviting public feedback on the draft to refine and expand upon the proposed guidelines.

“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, OpenAI explained in a blog post. “This is a continuation of our ongoing commitment to improve model behavior using human input, and complements our collective alignment work and broader systematic approach to model safety.”

Follow @voicebotaiFollow @erichschwartz

OpenAI Promises Creators New Tool to Control How Their Content Trains Generative AI Models

OpenAI Enhances Assistants API with Advanced File Management and Cost Control Features

OpenAI Enhances GPT-4 Turbo Model for API and ChatGPT, Including Visual Input