Obviously the scale of retrieval from memory is different, I'm not disputing that. The empirical question I was referring to is what percent of its training set the model is capable of reproducing. What looks like "large swaths" of text to us could be less than one in a million for all we know.

I'm skeptical that there is a single spectrum like you're describing. It's not well defined. Say a human and an LLM prove a new theorem independently (without external help, i.e. from their own neural weights and reasoning). How do we measure how much each of them copied from previous work, as opposed to having learned from or been influenced by it?