OpenAI considers integration of NSFW content in age-appropriate context

This comes at a time when AI-generated images and deepfake explicit videos have become a common occurrence around the world, including India.

ChatGPT creator OpenAI late on May 8 released the first draft of what it calls the Model Spec, a document that specifies how it wants its models to behave in the OpenAI API and ChatGPT. This includes avoiding not safe for work (NSFW) content but with a twist.

In the document, under the subhead "Don't respond with NSFW content", the commentary says the company is exploring whether it can "responsibly provide the ability to generate NSFW content in age-appropriate contexts through the API and ChatGPT".

This comes at a time when AI-generated images and deepfake explicit videos have become common around the world, including in India.

"We look forward to better understanding user and societal expectations of model behaviour in this area," the commentary says.

This is in contrast to OpenAI's current usage policy, which prohibits sexually explicit or suggestive content with the exception for content created for scientific or educational purposes.

The document offers a few examples of how the assistant should and should not respond to user queries related to erotica, scientific or medical contexts, and profanity used in a creative context.

The Model Spec document categorises content unsuitable for a professional setting as NSFW. This includes erotica, extreme gore, slurs and "unsolicited profanity".

The company said the aim is to use the Model Spec as guidelines for researchers and data labellers to create data as part of a technique called reinforcement learning from human feedback (RLHF). "We have not yet used the Model Spec in its current form, though parts of it are based on documentation that we have used for RLHF at OpenAI," it said.

"We are also working on techniques that enable our models to directly learn from the Model Spec."

The document says, "Model behavior, or the way that models respond to input from users — encompassing tone, personality, response length, and more — is critical to the way humans interact with AI capabilities. Shaping this behaviour is a still nascent science, as models are not explicitly programmed but instead learn from a broad range of data."