Abstract: Generative adversarial networks (GANs) have proven to be one of the most powerful methods for modeling complex probability distributions and it is of great interest whether they can be used in scientific applications. A major obstacle to doing so is that GANs typically favor samples with high likelihood, producing model distributions with lower entropy relative to the data. Here, I describe a new kind of adversarial objective in which two neural networks compete to maximize or minimize the entropy of a collection of output units and develop a training schedule in which a measure of the entropy of the model distribution is kept close to the entropy of the data distribution throughout training.