When AI art has no author: Study finds generated images often can’t be traced to training data

A study by MIT CSAIL researchers shows that as generative AI models are trained on massive datasets, an "attribution decay" occurs, making it impossible to trace generated images back to individual training data or specific artists. By using a novel "diffusion ensemble" architecture to efficiently test counterfactual scenarios, the researchers demonstrated that removing individual images often has zero impact on the final output. These findings complicate ongoing legal debates surrounding copyright infringement, fair use, and how to fairly compensate artists when AI-generated works cannot be linked to specific sources.

(Source：MIT News | Massachusetts Institute of Technology)