DCGAN is initialized with random weights, so a random code plugged in the network would produce a totally random picture. Even so, as you may think, the network has many parameters that we could tweak, as well as the target is to find a location of those parameters that makes samples produced from random codes look like the teaching information.