MIT Study Finds 24 AI Models Defy Image Attribution, Undercutting IP Theft Claims
Updated
Updated · Gizmodo · Aug 25
MIT Study Finds 24 AI Models Defy Image Attribution, Undercutting IP Theft Claims
1 articles · Updated · Gizmodo · Aug 25
Summary
Nature Communications published an MIT study showing diffusion-model outputs cannot be causally tied to specific training images, faces or artists once datasets grow large.
Two researchers built 24 custom models trained on a few hundred to hundreds of thousands of images, then removed individual images or segments to test whether outputs changed in a measurable “counterfactual radius.”
Across general-image, face-only and artist-specific experiments, larger training sets made attribution harder—a pattern the authors call “attribution decay”—and removing a source image often left the output unchanged.
That finding weakens artists’ efforts to prove a generated image copied a particular work, especially because commercial systems such as Midjourney and Stable Diffusion are far larger and likely even less attributable.
The paper broadens the legal and regulatory challenge around AI accountability, suggesting courts and policymakers may struggle to trace influence even when outputs closely resemble human-made art.