DCGAN is initialized with random weights, so a random code plugged in the network would make a totally random image. However, when you may think, the network has numerous parameters that we are able to tweak, and the target is to find a setting of such parameters that makes samples generated from random codes appear to be the education information.