They are just predicting the next token. In human text it's more common to talk to other people than a computer, so they end up talking to the computers like they were people.