DCGAN is initialized with random weights, so a random code plugged in the network would make a totally random image. Nevertheless, as you may think, the network has a lot of parameters that we could tweak, as well as the objective is to locate a placing of these parameters which makes samples produced from random codes seem like the training facts.