I see what you're saying, my issue with this is the product is (as I understand) no more than an amalgam of its inputs. I do understand the similarity to human artists, where one's art is building from reference (either directly/indirectly/cumulatively). The difference here for me is that current models, don't/can't comprehend the meaning behind the components of their construction. They also don't or aren't able to add any additional meaning to what they produce. I'm not sure that makes much sense. What I'm trying to communicate is more of a feeling behind the art, which is either really difficult to describe, or I lack the words. Maybe you can help with your own thoughts/corrections?
That second paragraph makes perfect sense, especially tying in to the first sentence of your first paragraph. I wonder if it might be possible to escape the necessity for human produced data for training? That would certainly alleviate a lot of my concerns with the tech, especially when talking local.
Discrub might be an option depending on access. It seems to work fine, but depending on the amount of data it is pretty slow. I've had it running since early this morning, about 12 hours now. Not sure I'd trust the extension with highly sensitive data, but it might be worth taking a look.