A key area of debate in generative AI has centred on where exactly the outputs from AI models come from. When an AI model creates a new painting, for example, is there a way to know which images from its training data gave rise to that output? When the resulting painting mimics a specific artist’s style, is it fair to claim that the artist is the creator of the new painting?

According to a new research paper by Zheng Dai and David K. Gifford of MIT, it will increasingly be difficult to do so as more training data is used to train generative AI models. The research centres on one type of generative AI model called diffusion models.. the type of technology that powers many generative image models. But it extends to other applications as well: diffusion models are also used to generate video, audio, protein modelling and therapeutic design.

The paper’s authors have termed what they found “attribution decay”. Simply put, as the training data increases for a generative AI model, the influence of each individual training sample (e.g. image, author or artist) on the generative AI model’s output diminishes. In fact, it can reach a point where the output of the generative AI model barely changes when a training sample is removed.