MIT finds AI images lose their source trail

MIT CSAIL researchers found that as diffusion-model training sets grow, the influence of any individual image on a generated output shrinks under an inverse power law—a phenomenon they call attribution decay. In the MIT study, ensembles trained on datasets ranging from 256 to more than 160,000 images produced outputs whose counterfactuals became harder to distinguish, after tests against 24 conventional diffusion models and 1,282 separately trained models. Creative-AI teams should not assume that better provenance tooling alone will settle whether an output copied a particular work; Claude’s watermark and C2PA approach is a useful complement, not proof of source attribution.