AI Art Is Untraceable to Its Training Data, MIT Study Says

MIT CSAIL researchers show that for diffusion-model image generators, attribution to individual training images is effectively impossible at scale: removing specific training images hardly alters outputs, a phenomenon they call attribution decay. The study finds outputs can resemble a human artist or artwork, yet it’s virtually impossible to prove which training data influenced the result, complicating IP lawsuits and signaling broader regulatory and accountability challenges for large commercial models.