This could actually be a good way to sidestep the training set copyright and access right issues. Copyright protection should solely encompass the expression of

This could actually be a good way to sidestep the training set copyright and access right issues. Copyright protection should solely encompass the expression of human generated content and not the underlying concepts. By training model B using the results generated by model A, the copyright of corpus_A (OpenAI RLHF dataset) remains safeguarded, as model B is never directly exposed to corpus_A, preventing it from duplicating the content verbatim. This process only transmits the concepts originating from corpus_A, which represents universal knowledge that cannot be claimed by any individual party.