MIT CSAIL researchers found that as diffusion models grow larger and are trained on more data, individual training examples have progressively less measurable influence on generated outputs, a phenomenon they call 'attribution decay.' Using an ablation method on 24 diffusion ensembles trained on datasets from 256 to over 160,000 images, lead author Zheng Dai and co-author David Gifford showed that removing specific artists' work from training data barely changed generated outputs at scale, making it 'practically impossible' to trace outputs back to specific training data. The findings raise questions about copyright, derivative works, fair use, and compensation for original creators, and come amid ongoing lawsuits against Stability AI and Midjourney.
Questions this post answers
Why can't AI image outputs be traced back to specific training images as diffusion models get bigger?
As diffusion models scale up in size and training data, the measurable influence of any single training example on generated outputs shrinks dramatically, a phenomenon MIT CSAIL researchers call attribution decay. Using ablation experiments across 24 diffusion ensembles trained on datasets from 256 to over 160,000 images, researchers found that removing individual artists' work from training barely changed what the model could still reproduce, making attribution to specific data points practically impossible at scale. Follow how AI copyright disputes evolve as attribution research reshapes fair-use arguments, tracked on daily.dev.
What did MIT researchers use to test whether AI models can be traced back to specific training images?
MIT CSAIL researchers used an ablation method on a 'diffusion ensemble' architecture made of many components trained on different data subsets, allowing components to be swapped out without full retraining. They trained 24 ensembles on datasets ranging from 256 to over 160,000 images pulled from sources like ArtBench, CIFAR-10, Fashion-MNIST, CelebA, and MetFaces to measure how much removing specific images changed outputs. Developers building or auditing generative models can track methodology like this via daily.dev.