MIT study shows AI art attribution fades as training data grows

In a new paper, MIT's Computer Science and Artificial Intelligence Laboratory built a generative AI using public-domain art from 744 creators and produced a single portrait. Researchers then removed each artist’s data one by one, regenerating the image each time; nearly all results were indistinguishable from the baseline, illustrating “attribution decay.” To achieve this without retraining the whole system, they partitioned the model into many small, independently trained components, later combined as a diffusion ensemble.

The study argues that repeated visual motifs across vast datasets mean no single work is essential for a given output. While the experiment focused on unprompted generation, experts warn that style-specific prompts may behave differently. Copyright scholars say the finding makes it harder to prove a specific work influenced an AI output, affecting both plaintiffs and AI firms in ongoing litigation.