For these purposes it doesn't matter if they are persons and they don't need to be anthropomorphized: it only matters that the LLMs can incorporate the data from person-generated works into their output, either to weight things or to be read by an actual human.
It actually matters quite a bit that they are not persons from the simple fact that LLM output cannot trivially be used as LLM training material without reducing the resulting models to eventual incoherence. There's a proof of this somewhere in the last year or two.
There isn't, today, a good filter for such input beyond that it came from a person or that it came from a probabilistic vector distance algorithm. Perhaps we'll have such qualification in the future to make the distinction in this context irrelevant.
