Researchers at MIT published a paper in Nature Communications showing that as diffusion models grow in size, it becomes increasingly difficult to trace any specific output back to particular training examples. The authors name this effect “attribution decay,” and they conclude that at sufficient scale attribution can become effectively “impossible.”
What the team did
According to the paper, the researchers removed the Mona Lisa and all known works of Leonardo da Vinci from the training set of a large generative diffusion model. Despite those removals, the model still reproduced the Mona Lisa and other works attributed to da Vinci. The result suggests that the model’s internal representations and generalization mechanisms can recreate material even when explicit source examples are absent from the training data.
Legal and industry implications
These findings have direct relevance to ongoing copyright litigation. For several years, artists have sued companies including Stability AI and Midjourney, and plaintiffs have sought access to the models’ training datasets on the premise that seeing the data would prove unauthorized copying. The MIT results indicate that even full access to training data may not provide the needed evidentiary link if the model is large enough.
The authors note that a model large enough to ‘‘absorb everything’’ may erase the ability to tie outputs to individual sources. From a legal perspective, that technical property could function as a practical defense for model developers, because the mathematical difficulty of attribution would make proving copying much harder.
Practical consequences
Scale has long been treated as a competitive advantage in the AI industry because larger models typically perform better. The new study suggests scale can also act as a legal shield. The researchers point out that the straightforward way to restore traceability would be to reduce model size, but that approach runs counter to current development trends that aim to increase model capacity and performance.
Summary
The Nature Communications paper demonstrates that growth in model scale can undermine the ability to attribute generated outputs to specific training sources. That technical phenomenon—attribution decay—means that scaling up models may provide developers with a de facto legal advantage by making it mathematically difficult to link outputs to particular training examples.
