LLMs didn't feel. They're a very complicated graph that connects pieces of words and their probability to be near other pieces of words.

A graph which was trained off the dregs of Internet forums. A lot of people say shitty things in the Internet, so those responses are heavily weighted in it's probability matrix