MIT research finds large training data pools make individual AI training samples unattributable

A MIT CSAIL paper, ‘Outputs of Generative Diffusion Models are Often Unattributable,’ shows that as training datasets grow, the influence of any single image or artist approaches zero. Using ablation methods, researchers found that removing individual training samples produces negligible change in outputs of large diffusion models. The findings weaken the ability to prove specific works were essential to training and may strengthen legal defences for image-generation systems trained at frontier scale.