Another (semi serious) possible explanation of this is that LLMs were trained on a huge volume of text from Reddit. I wonder if you'd see similar inhuman-looking behavior if you did an iterated prisoners dilemma with a group of Reddit power-users as the participants?