Analysis
MIT researchers Zheng Dai and David K. Gifford found that as generative AI models scale up, the link between their outputs and any specific piece of training data effectively dissolves -- a phenomenon the pair term 'attribution decay,' according to The Register. The finding cuts against a working assumption behind much of this year's AI copyright litigation: that a model's ability to reproduce a protected work is evidence it was trained on that specific work.
The team's methodology was ablation testing -- deliberately removing specific training examples, including Leonardo da Vinci works such as the Mona Lisa, and then checking whether a model could still reproduce them. For smaller models trained on narrower datasets, removing a source typically degrades or eliminates the model's ability to reproduce it, which is the intuitive result most copyright arguments rely on. But for sufficiently large diffusion models -- the class of systems underlying tools like Midjourney and Stable Diffusion -- the researchers found reproduction capability often survived the removal of the specific source entirely, because the model had absorbed the style and composition from the broader statistical patterns across its full dataset rather than from any single traceable example.
Why the legal system built its case on the wrong assumption
"If those outputs have nothing to do with any individual piece of training data, that raises questions about fair use," Gifford said -- a framing that cuts both directions. It could support AI companies' fair-use defenses, since an output that cannot be tied to a specific copyrighted input is harder to prosecute as a derivative reproduction of that input. But it also undercuts the opposite argument AI labs have sometimes made in their own defense, that removing objectionable content from training data reliably scrubs it from a model's outputs -- attribution decay suggests that once a large enough model has learned a style or pattern, deleting the original source and retraining does not necessarily erase the model's ability to reproduce it.
The practical stakes are immediate. Regulatory proposals in the U.S. and EU that would require AI labs to disclose or license specific training sources assume attribution is technically feasible at scale; this research suggests that assumption breaks down precisely in the largest, most commercially significant models, where proving or disproving that a specific output derives from a specific input becomes an open technical question rather than a settled forensic exercise. For AI companies currently defending training-data lawsuits, and for plaintiffs trying to prove infringement, the finding shifts the fight toward statistical and probabilistic arguments about likelihood of derivation -- a far messier standard than the direct-copying framework courts have used so far.
