‘Attribution decay’ complicates the picture of AI-generated images, scientists find

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory, Zheng Dai and David K. Gifford, reported in Nature Communications on 18 August that large AI image models can exhibit “attribution decay,” meaning removing specific training items may not meaningfully change generated outputs. Using a method that removes a single piece of training data and regenerates an image, they found that in very large datasets no single item can be shown to have decisive influence, leading to “unattributability.” The finding complicates arguments that any given copyrighted image in a training set directly causes infringement in an output, though it does not eliminate the possibility of direct copying when models are prompted to replicate specific works. Dai suggested that, paradoxically, the more extensive the training set, the harder it may be to attribute an output to any one source image.

Read the full article at The Art Newspaper - International art news and events