I read this article. There is an answer: #residuals. Interestingly, everytime a #BayesianPrior in the LLM is accessed, it's weight should be tied to the residual the copyright holder gets paid. #DigitalCopyrightManagement https://www.theguardian.com/technology/2024/jan/08/ai-tools-chatgpt-copyrighted-material-openai