MIT researchers find AI outputs often cannot be traced to specific training data

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have identified a phenomenon termed ‘attribution decay’ in large-scale generative AI models. The study suggests that as training datasets grow, the influence of any single piece of data on the model's output diminishes to the point of being unmeasurable.

According to the findings, removing a specific image, the entire body of work by a single artist, or all photographs of a specific person from a massive training set often results in no appreciable change to the model's generated output. Lead author Zheng Dai noted that if removing a piece of data does not change the output, it is difficult to argue that the specific data point was responsible for that output.

These results present significant challenges for copyright litigation and regulatory efforts. Many legal disputes regarding AI rely on proving a direct connection between protected source material and the generated content. If the influence of a single work is spread across countless other examples in a massive dataset, creators may face increased difficulty in proving that a model has copied their specific expression.