Researchers found that large-scale AI generative models demonstrate attribution decay, meaning individual training images often have no measurable impact on specific outputs. This finding suggests that traditional copyright attribution may not apply to models trained on vast, monolithic datasets.