AI Copyright and Attribution Debate

MIT study on attribution decay finds large generative models do not copy individual works, using diffusion ensemble method. Challenges assumptions about output traceability, with major implications for artist compensation, model licensing, and fair use. Generated images cannot be reliably attributed to specific training data. New development: Anthropic announced watermarking Claude text for EU AI Act compliance – practical guide explains secret key, no quality impact, and real agency transparency stance. This adds a compliance layer to the attribution debate.