🤖 MIT CSAIL: Generated images often have no specific source
Zheng Dai and David Gifford described "attribution decay" in Nature Communications: the more training data a diffusion model has, the weaker the connection between an image and specific examples in the dataset. A precise check was provided by an ensemble of small models on data shards — it is enough to disable the components that saw the example.
🌍 This harms copyright disputes: if removing a painting or all photos of a person does not change the model's output, it is difficult to prove derivation from a specific author. Entire data shards are subject to attribution, but this is not a legal conclusion.
👤 For engineers and MLOps teams: training on shards makes "forgetting" data precise and simplifies handling deletion requests. The cost is the quality and cost of an ensemble instead of a single model.
Source 1: https://news.mit.edu/2026/when-ai-art-has-no-author-generated-images-often-cant-be-traced-to-training-data-0818