Generative adversarial networks, or GANs, are widely used to create amazing “fake” images and realistic, synthetic training data. And yet, despite their name, mainstream GANs generate only the examples that are easiest to find, rather than the promised adversaries, which would be more diverse, more challenging and more useful. Careful re-examination and rethinking of the strategy underlying GANs leads to a novel refinement of GANs, which offers much better enrichment of datasets and a wider variety of generated images.