Since the launch of ChatGPT, OpenAI has been at the forefront of generative AI, setting multiple precedents for the future of AI development. In order to further establish transparency with the public, OpenAI has shared a new document that allows users to gain a deeper understanding of its AI models.

On Wednesday, OpenAI released the first draft of its "Model Specification," which is a document outlining how OpenAI intends its models to behave in the OpenAI API and ChatGPT. The "Model Specification" also aims to provide insights into how OpenAI makes decisions about model behavior.

Model behavior refers to how the AI model reacts to user input, including tone, response length, and other features. OpenAI explains that AI models are not explicitly programmed but rather learn from data, making shaping their behavior a "nascent science" that requires consideration of many subtle differences.

The "Model Specification" details how OpenAI tackles the complex task of shaping model behavior, including the experiences and ongoing research on the subject. OpenAI explains that it has not yet used the "Model Specification" in its "current form" but is researching techniques to enable its models to learn directly from the specification.

In the document, OpenAI breaks down three principles of model specification:

1. Goals, including assisting developers and end-users for the benefit of humanity;

2. Rules;

3. Default behavior, which refers to guiding principles consistent with the goals and rules.

The company states that it is sharing the model specification with the public as a guide for reinforcement learning researchers and AI trainers based on human feedback. By sharing this document, OpenAI aims to increase transparency regarding its models and gather feedback from all stakeholders on its approach.

The model specification feedback form is open until May 22nd. All stakeholders are encouraged to fill out the form, regardless of their technical expertise.

The model specification aligns with OpenAI's ongoing efforts to deepen trust with its stakeholders regarding its models. Just this week, OpenAI released a new tool that can detect AI images generated using its own DALL-E 3 tool.