Empirically, however, LLMs don't strip out the author: the big models know a lot about what I've written even with search disabled. Ex: https://claude.ai/share/8cbcdf88-a360-421a-8c06-ae7b7992e866
(I was pointing out that "they'll train on a version that strips out you as the author" seems to be is incorrect about how training works)
My experience being on the searching end is that these things are terrible about attribution of where they find anything. Which has bad consequences not just for authorship, but for correctness (which is the usual reason I'm poking at them -- they're being wrong again). This makes a lot of sense when you consider the massive, massive compression that's got to occur during training, but it's still frustrating.
The last one is the worry.
The middle is... where we all start.
The first is not a bad place to be, all things considered!
Are you actually saying you'd be OK with that?