Generative Adversarial Networks, here after referred to as GANs, are a kind of neural net where a generative neural net competes against a discriminating neural net. The goal of the generator is to create convincing fakes (images, sounds, what have you) and the goal of the discriminator is to select only the real images and discard the fakes. Put the two in competition over a set of millions of actual images of a flower. The generator throws out random visual static and after a few dozen tries the discriminator figures out the difference between random visual static and pictures of flowers. The generator steps up its game and starts putting in patches of one color and in a few hundred cases the discriminator has figured out that flowers are more than just blobs of color. Radial symmetry, petals, leaves, light and shadow, etc. All of the parts that make an image believable and lend it verisimilitude are added.

Does this work? Does the generator create convincing fakes? Which of these faces are real? Can you spot the fake(s)? Take your time. It's not even that hard if you know what to look for. All? None? Time is up! Answer here. Oh, and you don't get any points for figuring it out by reading the link text. There are tells but you have to know what you're looking for. Here's an AI that's constantly dreaming of human faces. It's just off enough that it makes good practice. It's an open question how far generative artificial intelligence can go but at the extreme we may see all art forms become AI wrangling.

IRON NODER XIV: THE RETURN OF THE IRON NODER