DCGAN is initialized with random weights, so a random code plugged in the network would make a completely random image. On the other hand, while you might imagine, the network has countless parameters that we can tweak, plus the purpose is to find a setting of such parameters that makes samples created from random codes appear like the instruction