These are because of post-training. You have to give it such directives in post-training to correct the biases they bring in from scraping the whole internet (and other datasets like books, etc.) for data