DCGAN is initialized with random weights, so a random code plugged to the network would make a totally random image. Even so, while you might imagine, the network has millions of parameters that we can easily tweak, as well as purpose is to find a location of those parameters that makes samples created from random codes appear like the instruction