>It doesn’t really matter who wrote it, human or LLM. The only responsible party is the human and the human is 100% responsible.

Yes it does.

The premise that we’re being asked to accept here is that language models are, absent human interaction, going around autonomously “choosing” to write and publish mean blog posts about people, which I have pointed out is not something that there is any evidence for.

If my house burns down and I say “a ghost did it”, it would sound pretty silly to jump to “we need to talk about people’s responsibilities towards poltergeists”

I don’t get your analogy. If you paid the ghost $20/month for its services and configured that ghost to play with fire with no supervision, then it is 100% your responsibility that the house burned down.

The point is that if somebody says a ghost burned their house down it is much more likely that they are lying than it is that they have discovered objective evidence of the existence of ghosts.

Similarly, there is no actual evidence that a language model, absent any human intervention, chose to autonomously write and post a mean blog post. It is far more likely that a person got mad and wrote a mean blog post than it is that we have witnessed the birth of a whole new phenomenon that is somehow simultaneously completely emergent and also has only happened one time.