The tension between copyright and LLMs can never be resolved in a traditional matter, because the training process can't keep any correlation between the model's behaviors and the source data.

The only accurate source to site for a prompt response would be the entire training set.