Great article, its just i don’t get something. If we set the discriminator to be non-trainable, how does it know(learn) what is real and what is fake?!

I know that its supervised learning and we feed it {0,1} but what is the role of its weights?!