001 Crossbreeds

A.I. photography shows us a recognizable world, but viewed askew, as if what we see is made of alien substances. Yet, despite their strangeness and glitches, these A.I.-generated images are bursting with legibility because the raw materials are quarried from our great cultural image banks.

The image generators train themselves on vast scrapes of internet photographs, text, images. With blissful lack of discernment, A.I. mines ‘straight’ photographs and fictions alike. The results it produces are admixtures, hybrids, proud crossbreeds. Consequently, A.I. imagery deftly tracks to and for between the apparently real and the discernably fabricated.

The results lead us, and perhaps the entire culture, in the same direction: to a place where fact and fiction blur, where truths and lies become indistinguishable, where we move toward mysterious new worlds like ships into deep space.

Jeff Alu. Artbot #1 – Fictitious First Attempts from the project Traffic Signals of Rebellion. Courtesy of the artist.

002 How It Works

Think of a game with two players. Player 1 is inhabiting the role of a “generator.” They create an image and pass it on to Player 2. Player 2 is assigned the role of “discriminator”—their job is to figure out whether the image created by Player 1 is real or not. Player 2, the “discriminator,” has a large database of “real” images next to them so they can easily check the images created by Player 1 against that data set. The goal of this game is for Player 1 to produce an image that Player 2 deems as “real,” or at the very least can’t discern as “fake.”

Here’s the interesting bit—whenever an image is deemed as “fake” by Player 2, Player 1 will receive some feedback from Player 2 as to why. Player 1 will then incorporate those notes into making a better image. This game repeats for as many rounds as needed until Player 1, the generator, produces images that Player 2, the discriminator, can’t tell apart from those in their dataset.

The above is an extremely reductive breakdown of how generative adversarial networks (GANs) work, with the two adversaries being the “generator” and “discriminator.” Many platforms, such as Midjourney, DALL-E, and now even Adobe Photoshop, rely on such adversarial networks to make their images. Many of those platforms incorporate a text-to-image feature too. In those instances of text-to-image generation, the system is trained on the relationship between text and image by playing a similar game.

Sam Finn. Pool. Courtesy of the artist.

003 Rorschach

A.I. photography is a Rorschach machine. We project ourselves onto these images. In them, we see ourselves, our world, and, particularly, a glimpse of the future.

We live in an age of uncertainty, not from a lack of information, but from too much. Every faction shops for what it believes—constructing insular worldviews and becoming easy marks for grifters and conmen. Common ground falls away, civil society seems imperiled. Many are confused by the present, fearful of the future. We are beset by questions. Answers seem urgent and we look for them everywhere, feel for them in the wind.

In this context, A.I. image generation becomes a pattern of tea leaves in which we try to read our destiny. We look to it for foundational questions. Who are we? What are we becoming? Where are we going? What are we doing to the world?  A.I. becomes an uneasy toss of predictive dice. We read the returns as portents, omens, and much of what appears depicts disorientation, darkness, and descent.

Ben Millar Cole. Untitled from the project Hand Device. Courtesy of the artist.

004 Nonhumanness

The artificial intelligence software used by these artists are, in the main, text-to-image models: DALL-E 2, Midjourney, Stable Diffusion, others. Artists enter a statement, provide a description, supply a string of words. What will the A.I. image generator produce? The results are a surprise.

The text inputs are essentially conjectural. They are questions, inherently speculative. If I input this description, this request, what image-answer will appear? What will A.I. conjure and what will it mean?

We trust A.I.’s inscrutable hands-off-ness, its automated nonhumanness. A.I.’s enigmatic operations scour worldwide seas of data. A.I. reveals hidden patterns, uncovers deep, uncomfortable truths. Its very automated quality, like a trawl net in the ocean of the digital image world, brings up remarkable creatures and unsettling monsters. A.I. is our mirror and we believe the mirror.

Diana Nicholette Jeon. Flood Unleashed from the series Unleashed. Courtesy of the artist.

005 Photography? Yes.

A.I. photographs are deeply photographic. They are meta-photography—photographs built by powerful processors and computational algorithms swirling together ripples left by of millions of previous photographs. They are the offspring of our culture’s photographic promiscuity mated with immense computing power.

In fact, A.I. photography reflects and pays homage to 170 years of photographic practice—the images we use to define and describe our world. It is a voracious system that devours all previous photographs and shares them back. It is thus utterly photographic and immensely attentive, inclusive, and generous.  How can an image summoned into being through the digital distillation of millions of photographs not be photographic?

Roope Rainisto. Ego Never Sleeps from the project Life in West America. Courtesy of the artist.

006 Photography? No.

Images generated by A.I. are not photographs. To be literal, “photograph” is a word whose Greek roots mean “written in light.” Light is required. Let us take as an example the black-and-white photograph titled PSEUDOMNESIA: The Electrician. The photo was top winner of creative photography in the Sony World Photography Awards, awarded April 13, 2023. The photo shows a melancholy older woman, eyes downcast, leaning in from behind and grasping the shoulders of a younger woman. It looks to be a vintage wet-plate collodion photograph. “Haunting,” states the press release, “reminiscent of the visual language of 1940s family portraits.”

The winning photograph, however, emerged from A.I. image generator DALL-E 2, as prompted by Berlin-based artist Boris Eldagsen. The photography community exploded in consternation and outrage—not a photograph at all. Light was not involved, consequently this is not a photograph. Eldagsen declined the honor. “A.I. images and photography should not compete with each other in an award like this,” he wrote. “They are different entities. AI is not photography. Therefore, I will not accept the award.”

Karchi Perlmann. From the project Remembering. Courtesy of the artist.

007 Threat

Artificial intelligence has moved quickly from mirage to marvel to threat. Some credible experts now view it as the classic science fiction cliché—an avenging technology that rises to destroy its maker. Congress staggers toward regulation. The word “apocalypse” crops up.

The artists in this exhibition have informed attitudes toward the new technology. In the end, though, they have decided to use A.I. image generators as what they are: a new tool. The exhibition itself takes a similar stance: let’s look at this explosion of possibilities, but do so with clear eyes.

A.I. image generation poses ethical problems that fall short of apocalypse. Here are several. Text-to-image models are trained on huge datasets of text/image pairs scraped from the web. Some characterize this as industrial level plagiarism. Litigation against Stable Diffusion alleges that the image generation tool trained its model on “millions of images protected by copyright.” Then there is the threat of displacement. Molly Crabapple, artist and activist, states A.I. image generators “are built for the purpose of deskilling, disempowering and replacing real, human artists.”

Boris Eldagsen. Hunger # 17. Courtesy of the artist.

008 Recursive

“So we beat on, boats against the current, borne back ceaselessly into the past.” – from a famous book by a famous author

The relentless recycling of trends, tropes, and images is evident in the workings of modern mass culture. The past can be a source of comfort and escape; it’s a “known” place and space. Many movies, shows, and records are deeply engrained into our personal memories and biographies all the while serving as an overarching connection to culture at large. The past is extremely profitable too. The success of past media blockbusters serves as a template for future development, causing a cascade of countless sequels, reboots, and actors brought back to life through CGI.

Enter, then, AI image systems. Trained upon our vast, accumulated reservoirs of cultural imagery, these systems are machines put to the task of magnifying our recursive tendencies. We may find ourselves trapped in a cultural Möbius strip, where rather than driving us forward, AI reinforces a perpetual feedback loop of the past. A feedback loop that often declaws many of the radical, experimental, and avant-garde movements that challenge (or resist assimilation into) the dominant cultural hegemony of a given moment. The question we are confronted with then is whether we are consigning ourselves to a cultural future that is forever trapped in a skewed and simplified version of its past. The words of the late scholar, author, and critic Mark Fisher’s come to mind: “Those who can’t remember the past are condemned to have it resold to them forever.”

DALL-E 3 / OpenAI