New Developments in Addressing Copyright and Compensation Issues in Generative AI

Researchers at Carnegie Mellon University have recently conducted a study on copyright and compensation issues in generative AI models.

The Generative Intelligence Lab at the School of Computer Science collaborated with Adobe Research and the University of California, Berkeley to develop two algorithms that make significant progress in addressing these issues in generative AI models. The first algorithm prevents these models from generating copyrighted materials, while the second algorithm develops a way to compensate human creators when the models use their works to generate images.

Powerful tools like DALL-E 2, Midjourney, and Stable Diffusion, which are image generation models, create realistic visual content from simple textual descriptions. Behind the scenes, these models are trained on millions to billions of internet images, including potentially copyrighted materials, licensed images, and personal photos.

"As researchers in this field, we have a responsibility to address the social issues associated with it. Technological solutions are just one aspect. We also need to do more work in legislation and regulation of artificial intelligence," said Junyan Zhu, Assistant Professor at the School of Computer Science and head of the Generative Intelligence Lab.

The research team will present two papers at the International Conference on Computer Vision in October 2023.

The first paper, "Removing Concepts in Text-to-Image Diffusion Models," helps AI generation models avoid generating specific copyrighted images or styles.

For example, if you ask an AI program to generate a painting by a living artist, it will generate an image closely resembling the artist's style. The algorithm proposed by Carnegie Mellon University researchers aims to prevent this from happening and prompts the AI model to generate a generic artwork.

"We can offer this as an option for artists to opt out of AI model usage at any point in time. It provides more control and freedom for individuals and companies who do not want their images to be used," said Nupur Kumari, lead author of the paper and a Ph.D. student in robotics.

The second paper, "Evaluating Data Attribution in Text-to-Image Models," develops a method to compensate individuals and companies who contribute data to train AI. The algorithm attempts to determine the contribution of each training image to the generated image and can be used to fairly allocate payments to owners of copyrighted images in the AI database.

For example, if you ask an AI model to generate an image of a watercolor painting, the final image will be influenced by artists skilled in watercolor painting. The goal of this new algorithm is to quantify the contribution of each artist to this new synthesized artwork.

"Our goal is to answer the question, 'Which set of images influenced the synthesized image?' We hope to use this algorithm to allocate rights to data contributors. The ultimate goal is to fairly compensate data owners who contribute to the creation of generative AI images," said Sheng-Yu Wang, lead author of the paper and a Ph.D. student in robotics.

These new algorithms are still in the early stages of development, and the authors acknowledge that there are still many unanswered questions. For example, it is unclear whether copyrighted content has been completely removed or just hidden somewhere, and further research is needed to explain how the attribution algorithm evaluates the impact of each training image.

Despite some unanswered questions, these new algorithms take the first step towards addressing copyright issues on generative AI platforms and compensating individuals and companies who contribute to AI image creation.