DCGAN is initialized with random weights, so a random code plugged to the network would create a very random image. Even so, as you might imagine, the network has millions of parameters that we can easily tweak, plus the target is to find a setting of these parameters that makes samples produced from random codes seem like the coaching details.