AI Models' Source Attribution Becomes Increasingly Difficult as They Grow

MIT researchers have found that attributing AI model output to specific training data becomes more challenging as models grow in size. The study, titled 'Outputs of Generative Diffusion Models are Often Unattributable,' will be published in Nature Communications on Tuesday. The authors, Zheng Dai and David K Gifford, note that diffusion models like Midjourney and Stable Diffusion have become widely used for generating artifacts such as images, videos, and audio. The researchers tested this by removing specific training data through a process called ablation, and the result showed that for very large models, outputs could still be reproduced even when specific images or works were removed from the training dataset.