The generator implicitly models a joint probability of data by being a generative process that one can draw samples from. GAN training (at least under certain simplifying assumptions) minimizes the JS divergence between the generator distribution and the data distribution.