Research in computer graphics has been in pursuit of realistic image generation for a long time. Recent advances in machine learning with deep generative models have shown increasing success of closing the realism gap by using data-driven and learned components. There is an increasing concern that real and fake images will become more and more difficult to tell apart. We take a first step towards this larger research challenge by asking the question if and to what extend a generated fake image can be attribute to a particular Generative Adversarial Networks (GANs) of a certain architecture and trained with particular data and random seed. Our analysis shows single samples from GANs carry highly characteristic fingerprints which make attribution of images to GANs possible. Surprisingly, this is even possible for GANs with same architecture and same training that only differ by the training seed.