The GAN seems to be able to conjure particular latent anxieties based on our perceptual position. For example, if you ask the BigGAN to generate objects in the sky it seems to possess a bias towards violence. To the GAN birds look like far off fighter jets dropping bombs. The BigGAN seems to perform the individual looking up to the sky under a culture of the violence of optics, a world where a camera sight drone can reign terror from thousands of feet in the sky. In the BigGAN objects in the sky are violent, disembodied, and militaristic. 

​

If the GAN is a camera, then who stands behind it? How does the GAN perform us, the photographer? Behind every BigGAN image there is a photographer, a ground-mounted body straining to point a heavy lens towards the sky and photograph a bird, or a sweaty finger in balmy heat photographing a seafood platter. Because the BigGAN is trained on human images it generates photographs that have a photographer behind them, an implicit ghost in the inverse of every GAN image, a disembodied photographer. This disembodiment is different from the disembodiment inherent to the driverless car camera or surveillance camera which whilst not anthropocentric, do possess physical (plastic or metal) bodies that dictate the kinds of images they produce. GAN images in their human perspective point to an actual human limited photographer. An averaged idea of a photographer who is nevertheless limited by a body and its interaction with a camera, we cannot stretch our arms into the sky to photograph birds, for example.(1) The GAN is disembodied in that its body is an abstract, an average. Who then is this median ghost behind the image?

When the BigGAN turns on us, it does so relatively unsuccessfully. Within photographic frames, to the averaging eye of the GAN, bodies are unruly and patternless objects. We hold the camera on our physical plane meaning that when we are photographed, we are infinitely variable. Unlike our bodies birds can be successfully averaged because, to the camera clutching person, they are all above and as such there is uniformity in a photograph of a bird in flight. We are amalgamated into strange meshes of flesh. Is this how the GAN sees us? Is this how it imagines the person behind its millions of photographs?

 

This depends on how the GAN has been trained. Some GANs, like StyleGAN, can be coaxed to see us in a kinder light. StyleGAN can generate faces that are indistinguishable from real people. To the GAN trained on the right dataset, a dataset exclusively of faces, facial features are a very stable category (hair was in-fact its most significant stumbling block). 

 

Images generated by the StyleGAN are the ‘last people’; the last photographs of people that ever need to exist.(2) StyleGAN generations are so convincing that they subsume all other photographs. A profile photograph of a person looking into a camera can now never again be assumed to hold any link to a real face. The StyleGAN is so effective at generating faces that new face generation GANs no longer need to be trained with real human human images. Rather, images generated by a StyleGAN can be used to train new GANs. The StyleGAN doesn't even need to learn from our images anymore. 

 

When the GAN’s internal processes succeed, it takes ownership over all photographic representation, all images become GAN images. In almost all generations by StyleGAN people are smiling. Sometimes, it generates those awkward half-smiles captured when a photograph is taken slightly too early. It captures the reflection of light in our irises and redeye. These reflections make the StyleGAN so convincing, not by generating ‘new people’ or other representations of bodies; drawings, diagrams, or measurement of weight and height, but through generating photographs of people. And don’t we privilege the truth of photographs above all else? In practice, the photograph has subsumed its representational function and exists now as the thing itself, our own faces becoming referent to images of our faces.