MIT study finds many AI images cannot be traced to training data

MIT CSAIL researchers identified a phenomenon called attribution decay, in which individual training examples matter less to the output of a generative model as the training dataset grows. At sufficiently large scale, removing a single image, all works by a particular artist, or every photograph of a specific person often does not appreciably change the generated image.

The team built a diffusion ensemble to test that counterfactual directly. Instead of retraining a model from scratch after deleting each example, the architecture uses smaller components trained on different slices of data, allowing researchers to switch off the parts that saw a given input and observe what the model would have produced without it.

The findings raise questions for copyright, licensing, privacy, and responsibility in generative media. MIT professor David Gifford said the work suggests some model outputs are not simply copies of training data and could help companies demonstrate that generated images are not derivatives of individual works or people. Whether the same attribution decay applies to large language models remains unresolved.