MIT Study Questions Traceability of AI-Generated Images to Specific Sources
New MIT research suggests that tracing the origin of images generated by AI models trained on large datasets to individual training data pieces is difficult. This finding could impact copyright lawsuits.
A new study from the Massachusetts Institute of Technology (MIT) challenges existing notions of copyright for AI-generated imagery. Researchers found that as the training datasets for AI models grow larger, it becomes increasingly difficult to attribute generated images to specific original works.
MIT researchers Zheng Dai and David Gifford discovered that the more extensive the training dataset, the harder it is to prove that any single piece of data significantly influenced the final output. They observed that removing an individual image from the training data often had little to no discernible impact on the model's newly generated images. The researchers term this phenomenon "attribution decay."
This finding could complicate legal cases where AI is accused of copying artists' work. While an AI-generated image might resemble a specific artist's style, this new research suggests there may not be a direct causal link connecting the created image to a particular piece within the training data.
The researchers hypothesize this occurs because large datasets contain substantial visual redundancy. Many different images share overlapping features, meaning no single image is irreplaceable for the "DNA" of an AI-generated image. While presenting this as their best explanation, they emphasize it is currently a conjecture rather than a directly proven fact.