I do have an hypothesis: If GPT-4 was trained with a substantial amount of internet conversations it could be because it's very uncommon for people to respond to questions on forums with "No sorry I don't know", instead the question/thread would just die without answers.