OpenAI has been a leader in the generative AI space since its launch of ChatGPT, and with that role has come the responsibility of setting precedents for future AI developments. To that end — and further foster transparency with the public — OpenAI shared a new document that gives users deeper insight into its AI models.
Also: Is this mystery chatbot really GPT-4.5 in disguise? Here’s how to see for yourself
On Wednesday, OpenAI made public the first draft of its Model Spec, a document that outlines how the company wants its models to behave in OpenAI API and ChatGPT and provides insight into how the company makes decisions about its models’ behaviors.
To deepen the public conversation about how AI models should behave, we’re sharing our Model Spec — our approach to shaping desired model behavior. https://t.co/RJBRwrcTtQ
— OpenAI (@OpenAI) May 8, 2024
Model behavior refers to how AI models respond to user inputs, including characteristics such as tone, response length, and more. OpenAI explains that AI models are not explicitly programmed but rather learn from data; therefore, shaping their behavior remains a “nascent science” with many nuances to account for.
The Model Spec includes documentation of how OpenAI approaches the complex task of shaping model behavior, including experiences and ongoing research regarding the topic. OpenAI explains that it has yet to use the Model Spec “in its current form” but is working on techniques that will allow its models to learn directly from the spec.
Within the document, OpenAI breaks down the three Model Spec types of principles: objectives, which include assisting the developer and end user and benefiting humanity; rules; and default behavior, which refers to guidelines consistent with the objectives and rules.
The company says it is sharing the Model Spec with the public to serve as a guideline for researchers and AI trainers who are working with reinforcement learning from human feedback. By sharing the document, OpenAI seeks to increase transparency about its models with its audience and collect all stakeholder feedback regarding its approach.
Also: Google was right to be worried: OpenAI reportedly wants to enter the search market
The Model Spec Feedback form is open until May 22. All stakeholders are encouraged to complete the form, regardless of technical expertise.
This effort fits into OpenAI’s ongoing efforts to deepen trust with its stakeholders regarding its models. Just this week, OpenAI released a new tool that can detect its own DALL-E 3-generated AI images.