An MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) report dated Aug. 18, 2026, described research finding that in generative diffusion models trained on large datasets, the link between individual training images and outputs can disappear. The researchers call this "attribution decay": as training data grows, removing any single image, or every image by one artist, often leaves a generated sample unchanged. The work appeared in an open-access paper in Nature Communications, according to the report.

To test the idea exactly, the team built a "diffusion ensemble" of smaller components, each trained on a different slice of data, letting them switch off parts instead of retraining. In comparisons with 24 conventional diffusion models trained on the same data, image quality was broadly similar by standard measures, and the ensembles scaled better with more data, the report says. Across 24 ensembles trained on datasets from 256 images to more than 160,000, drawn from seven public collections, the counterfactual radius shrank along an inverse power law as training sets grew.

The researchers said the finding may bear on copyright questions but noted whether it holds for large language models remains open. The work was supported by Schmidt Futures.