Unpacking the bias of large language models