MIT researchers identify attribution decay in generative AI models

Researchers at MIT have identified a phenomenon called ‘attribution decay’ in generative AI models. The study suggests that as training datasets grow larger, it becomes increasingly difficult to trace a specific AI-generated image back to any single piece of training data. At a certain scale, removing an individual image or even all works by a specific artist from the training set may result in no detectable change to the model’s output, making it nearly impossible to prove a specific image was responsible for a generated result.

To address this challenge, the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) developed a new architecture known as a ‘diffusion ensemble.’ Unlike traditional monolithic models, this system is composed of multiple smaller components, each trained on different subsets of data. This allows researchers to test what an AI would produce without a specific image simply by switching off the components that viewed that data, rather than retraining the entire model from scratch.

In testing, the diffusion ensemble produced image quality comparable to 24 conventional diffusion models. The researchers noted that while the ensemble performs poorly with small datasets, it scales more efficiently than standard models as the amount of training data increases.