the artificial archive dreams itself
Nearly a century after Carter took his photograph, I am sitting here, trying to see if DALL-E, an AI text-to-image system, might generate something similar. I start by entering the archival metadata and some basic description into the prompt bar: 1936 farm security administration photograph, “Dwelling, uninhabited, on proposed wild life area near Easton, Pennsylvania,” taken from inside of room, there is an open door in the center of the photograph . . . The program runs for several seconds. Doors and windows appear in black and white, next to or nestled in each other, looking into messy rooms and out onto barren worlds or blackness. I tweak my instructions, often in vain, trying to better approximate the old photograph. But it’s too late. The old photograph was just a beginning. I am passing through a tunnel of open and imaginary doors.
Describing something in the process of its undoing is surprisingly difficult. How do you name something already halfway gone? In Carter’s photo, what should I call the black rectangle leaning against the wall, the strips of paint or wallpaper on the floor? In the computer-generated images, names are even less apparent or useful. Doors don’t fit their jambs, walls lean precariously, slats are rough-hewn and uneven. The rooms are splotched and smudged, as if by charcoal or fingerprints. One would expect this on any old photo’s surface, but here, such wear is part of the architecture itself.
For this image-making machine, the texture of the past is the past itself, a result of the millions of old photographs this AI system was trained on, all those digitized copies of timeworn photographs. But it is startling to think that these brand-new images are more thoroughly warped with time than anything in the actual archive. They turn history into haunting, a kind of representation without depiction. Each blur a small collage of damage, each door a million doors.
Surveying these strange rooms, they do not look like collages at all. They look like photographs. Are they? There was no viewfinder, no shutter release, no man in a crumbling room. No government commissioned them; they capture no crisis. They were never taken, and yet they stand forth as photographs. It is impossible not to use the language of photography, not to say that the windows are overexposed or that the door is out of focus. It is impossible to see these for what they are: rearrangements of pixels based on textual inputs. They may not be convincing as photographs, but they are convincingly photographic. One starts to imagine having stood in each room with a camera, its shutter clicking away in the twilight.
I repeat my text-to-image quest with another one of Carter’s interiors, a rotting barn suffused with sunlight. The results range wildly, from entire barns to rafters to jumbled piles of wood—a result of my poor prompts, not nearly specific or thorough enough. Again the gulf between text and image yawns; again one wonders how to describe this rotting plank, that crossbeam, that floorboard. How to describe the quality of the light, not streaming in but seeping. But the AI system is not impressed by poetics and continues its infinite guessing. It will get it right not through attention but attrition.
Typing and retyping prompts begins to feel awfully Borgesian. I become aware, with each new set of not-quite-right images, that in theory there exists an arrangement of letters that would summon an exact arrangement of pixels to match the scan of Carter’s photo. I become aware of the immense but finite amount of pixels that make up any image, the ones and zeroes this machine keeps transmuting into picture. I find it all dizzying to contemplate, crouched in artifice’s sun-seeped crawlspace.
These new text-to-image generators, like DALL-E, tend to encourage users to collide different well-known styles, to make visual jokes and silly remixes. An object in the style of a famous painter (“A rubber duck as if painted by Van Gogh,” &c. ), for instance, or a famous character captured by a certain kind of camera (“Snow White and the seven dwarves caught on surveillance footage,” &c. ). It’s a funny gimmick that quickly becomes tedious. For these systems, it is a way of advertising the inexhaustibility of the system far more than the creativity of its users. The cheapening and vulgarizing of creativity, after all, is their business model: a million billion images that anyone can make—with very little effort, and for a fraction of what it’d cost to pay an artist.
When I try generating my non-photographs on Midjourney, another AI system, the open doors look comically stark in the torrent of other users’ content: clip art, smooth and 3D cartoons, superheroes, avatars, pseudo-photoshoots of video game characters, weird airbrushed faces and logos. This tends to be what people think of when they think of AI art: highly impressive images that are nonetheless needless and grotesque, over-iterated and uncanny.
There is something needless and grotesque about my little non-photographs, too, though I think these doors and barns constitute a different kind of seeing. They are perhaps best conceptualized, not as collages or even photographs, but as dreams. In a mechanical rather than mystical sense, the way that our dreams reassemble the noise of waking life into forms at once hyperreal and hallucinatory. It is not far-fetched to say that I am asking the photographic past to re-dream its own ruin. It should be no surprise that it does a decent job. It should also be no surprise that it tells us next to nothing about the past.
I do the same thing with more photographs from the Resettlement Administration and its successor, the Farm Security Administration. A collapsed porch (Porch of rehabilitation client, Boone County, Arkansas), a beehive-shaped furnace (Jefferson furnace, made iron for the Monitor in Civil War, Sackson, Ohio). Finally I try to regenerate a photo of a horse: Neglected horse owned by rehabilitation client, Jackson County, Ohio.
After looking at so many buildings, the horses are startling. Perhaps because, for something to be truly uncanny, it must be lifelike and not merely realistic. The most convincing AI barn is less uncanny than the least convincing AI horse. And many of my horses were unconvincing. It took much insistence to get them to face the right way, for instance. The AI system also kept interpreting blinders as other forms of blindness: eyelessness, deformity, cloth hoods, droopy manes.
Rearranging slats with a computer is a neater thought exercise than summoning the likenesses of things that never walked the earth; the horses seem more blasphemous than the barns. The director of the Farm Security Administration’s photograph program once explained the purpose of those images as “introducing America to Americans.” Ninety years hence, they continue to serve that purpose. But what are these AI images? What is their purpose? What do they do? What do they introduce?
They, too, introduce America: an unreal dominion of endless images, evicting history from its textures. There is no horse. Yet here it is. Its whitehot eye never blinks, never sees, never was.
ben tapeworm