MIT study finds AI-generated images often can't be traced to training data

MIT CSAIL researchers identify "attribution decay": the more data an image generator trains on, the less any single training image — or all images by one artist — affects outputs. Lead author Zheng Dai argues if deleting data doesn't change the output, it can't be attributed. David Gifford calls it the first method proving deleted inputs have zero influence.

Featured · Zheng Dai, David Gifford