MIT Researchers Find Attribution Decay Across 24 AI Models as Datasets Top 160,000 Images
Updated
Updated · MIT News · Aug 18
MIT Researchers Find Attribution Decay Across 24 AI Models as Datasets Top 160,000 Images
3 articles · Updated · MIT News · Aug 18
Summary
A Nature Communications paper from MIT CSAIL reports that in large diffusion models, removing any single training image—or even all images by one artist—often leaves generated outputs unchanged.
Using a custom “diffusion ensemble” architecture, the team created exact counterfactual models without full retraining and matched the image quality of 24 conventional diffusion models trained on the same data.
Across 24 ensembles trained on datasets ranging from 256 to more than 160,000 images, the maximum effect of any one training example shrank under an inverse power law, with the result surviving 1,282 brute-force retraining tests.
The finding could weaken attempts to tie specific AI outputs to specific source works, complicating copyright, fair-use and derivative-work claims while suggesting companies may need model designs that can prove outputs are unattributable.
The study covers diffusion models used in image and audiovisual generation; whether the same attribution decay holds for large language models remains unresolved.
Does this untraceable blending of data mean AI learns just like human artists, or is it the ultimate copyright loophole?
Could the discovery of AI attribution decay force the legal system to completely rewrite intellectual property laws for the modern era?
Attribution Decay in Generative AI: How Scaling Kills Traceability and Reshapes Copyright, Privacy, and Creator Rights
Overview
MIT researchers discovered 'attribution decay'—as generative AI models are trained on ever-larger datasets, the connection between any specific training image and the model’s output disappears. This happens because features and concepts become redundantly spread across thousands of data points, so removing an artist’s entire portfolio or a single image has no effect on the generated result. As a result, traditional methods for tracing outputs to their sources and compensating creators collapse, and AI developers can use extreme scaling as a shield against copyright liability. This forces courts and technologists to seek new ways to assess copying and raises major challenges for privacy and data governance.