There's no rule against learning from books in general. (Maybe only for some particular books.) There's no rule against using tools to read books (it's okay to wear glasses)

The only deontological argument I can see against training LLMs on copyrighted data is that some people think it's morally and legally wrong to make derivative works (such as fanfiction) without permission, and the weights of the LLM could be seen as a derivative work.

This may sound stupid but it's the same argument people use in favour of adblockers. When ads were first introduced to the internet, it was understood that different people could view the internet however they liked and they could choose a "user agent" to display content to their preference. So ads were just a nuisance but could be worked around easily -- it was the ad provider who was the fool. Now I often hear people say that ad blockers are unethical; they deprive content creators of their income, or they're deceptive, or criminal. This may indeed be true. But at no point in time did any moral rule suddenly change.

A website should not be able to dictate what application I use for passkeys, what hardware the passkey is stored on, or whether I'm allowed to duplicate my own key.

GP claims that it doesn't do anything until you tell it to do something. To me, this reads as GP implying that an agent won't do anything unexpected.

Humans are not autonomous: they're not even your employees until you hire them. (This is what I understand you're saying about AI.)