MIT Finds AI 'Attribution Decay' as Models Scale

MIT CSAIL researchers identified a phenomenon called "attribution decay" in large diffusion models ^[1]. The study found that outputs from these systems often cannot be traced back to a single image or artist from the training set ^[1][3]. This occurs because individual training examples lose measurable influence as models grow in size ^[2].

The findings complicate current efforts for copyright attribution and data governance ^[1]. When models scale, the contribution of specific creators becomes diluted within the vast dataset ^[2]. This shift poses significant challenges for legal frameworks designed to compensate or credit original authors ^[1][3].