Surrealism in the Age of AI

“This is not a pipe”: Why do AI Images Look Surreal?

By now you’re probably familiar with text-to-image programs like DALL-E, Midjourney, and Stable Diffusion, so you have a sense for the aesthetics of the images they produce: glitchy, plastic, uncanny. Even the “realistic” examples feel just a little surreal. There are practical reasons for this—and not all of them are technical.

Surrealists have long been fascinated by the slippery relationship between objects and the words we use to represent them. It’s the central idea behind one of the movement’s signature images, René Magritte’s 1929 painting The Treachery of Images. Famously, the artwork features just two basic components: a picture of a pipe and an accompanying caption that reads, “Ceci n’est pas une pipe” (“This is not a pipe”). Magritte’s gesture was both revolutionary and remarkably banal. Its wit works because we’re so used to negotiating the gap between signifier and signified that we almost forget the gap exists at all. But AI models, which “learn” through analyzing visual and textual data, are not so adept.

“Ingesting many different images of a ‘chair’ can allow a model to develop some kind of understanding of what a ‘chair’ might look like in different scenarios, but models don’t comprehend what a chair is as we might do,” explains the British-born, Berlin-based artist and theorist Mat Dryhurst. “In the course of generating the image, the system is trying to approximate something very new based purely on pixel data and concepts pieced together from that data.”

Dryhurst and his partner in life and art, the American composer and artist Holly Herndon, are among those driving the discourse at the edge of AI art. Their own work isn’t very surreal, but it is abstract, often because what they make is systems that make art. The duo has produced algorithm-aided electronic music, developed a program that allows singers to perform using a deepfake version of Herndon’s voice (they demonstrated it in a TED Talk, then used it to cover Dolly Parton’s “Jolene”), and helped launch a website that enables artists to remove their work from datasets used to train AI models. For the 2024 Whitney Biennial, they created a free app that produces ultra-exaggerated, “hairy mutant” pictures of Herndon. The goal is to produce enough of these user-generated pictures so that in the future, when commercialized text-to-image models create a portrait of her, the results will hew to the likeness she chose, not the aggregate one combed from the internet.

Herndon and Dryhurst’s Whitney contribution explores the fluidity of consent, identity, and intellectual property in the Web 3.0 world. But the project also points to one of the defining aspects of AI-produced images. Because these programs are synthesizing pictures from millions of jpegs scraped from online, the results are, by definition, amalgamated—and amalgamated images look unnatural. They’re like those composite faces used to illustrate the “averageness” theory of attractiveness. In the end, they all look like Jesus: a little familiar, kind of hot, completely unmemorable (also, far too often, inexplicably white). With all notable characteristics blurred to the median, their beauty is, paradoxically, the average kind.

The German artist Hito Steyerl uses the term “mean images” to describe these AI-generated composites. “They are after-images, burnt into screens and retinas long after their source has been erased,” she wrote in a New Left Review essay last year. “They perform a psychoanalysis without either psyche or analysis for an age of automation in which production is augmented by wholesale fabrication.” For Steyerl, “mean images are social dreams without sleep, processing society’s irrational functions to their logical conclusions. They are documentary expressions of society’s views of itself, seized through the chaotic capture and large-scale kidnapping of data.”

The Magnetic Fields: The Surrealist Logic of AI Models

Beyond these aesthetic considerations, AI image generators are, it turns out, surprisingly optimized for the exact brand of artmaking espoused by Breton in his manifesto—one that is absent of “control exercised by reason” and “exempt from any aesthetic or moral concern.” To liberate their creative impulses from the constraints of logic and received knowledge, he and other Surrealists relied on “automatic” exercises—that is, quick stints of writing or drawing without consideration for the coherence of what is being written or drawn. The Magnetic Fields, a 1920 book written by Breton and fellow French poet Philippe Soupault, was created via this technique. Credited as the first Surrealist literary work, it’s filled with arcane passages like this: “The corridors of the big hotels are empty and the cigar smoke is hiding. A man comes down the stairway and notices that it’s raining; the windows are white. We sense the presence of a dog lying near him. All possible obstacles are present.”

Automatism is light work for contemporary machine learning systems, which can process even the most complex computations faster than the time it takes to put pen to paper. For AI artists, the labor of creation is now outsourced to the machine; they need only to type a prompt and watch it work. “The rapid cultural rise of the ‘algorithm’ . . . testifies to the success of the Surrealist revolution that Breton never tired of promising,” wrote journalist Rob Horning in a 2022 Art in America essay. “AI cultivates and caters to our passivity, seeming to offer the fruits of creativity and self-examination without the effort and self-doubt.” In a way, Horning proposes, “the experience of AI in everyday life renders us default Surrealists, deferring to opaque automatic processes that no longer need be arduously evoked with Ouija-esque analog rituals.”

The Surrealists of the 1920s looked to the subconscious for an escape from the marching machine logic of the second industrial revolution. Breton once categorized his movement as a “violent reaction against the impoverishment and sterility of thought processes that resulted from centuries of rationalism.” But as Horning reminds us, AI does not provide the same escape; it is inextricable from the late capitalist forces that produced it. “Recuperated as AI, Surrealism provides the basis not for liberation but for further entrapment in existing cultural patterns reshuffled in novel ways, but not fundamentally changed,” he points out. “The idea of escaping from the control exercised by reason ends up being a way of fully submitting to a different form of programming, to what a machine learning model can produce and what algorithmic forms of control can induce.”

The Persistence of Memory: Do These Machines Dream?

Dreams, and to a lesser extent drugs, have become go-to metaphors for describing how AI image programs operate and the pictures they produce. It’s not a stretch to suggest that these associations can be traced back to one source: DeepDream, created by Google engineer Alexander Mordvintsev and released in 2015. The program visualizes pixel patterns learned by its artificial neural network, mimicking and exaggerating—to a psychedelic degree—the same process through which the human brain perceives the shapes of animals in clouds, say, or faces in wall outlets.

The project was a viral sensation when it debuted, and for good reason: Its creations are genuinely novel and mesmerizing. Pictures input into DeepDream come out looking like hallucinogenic fever dreams, kaleidoscopic mosaics of trippy textures and Freudian flashes.

Just as Dalí’s and Yves Tanguy’s symbological paintings of the 1920s, ’30s, and ’40s hinted at a realm of dreams lurking underneath our own world, DeepDream suggests that below the surface of every digital image is a roiling stew of other images, millions of them, all trying to bubble to the surface.

Because Mordvintsev’s project was many people’s introduction to AI outside the realm of speculative fiction, it would come to cast a long shadow over the ways in which we imagined, understood, and talked about the technology—perhaps to a misleading degree, since it isn’t really indicative of the way that machine learning models looks and work now. “I get it—the loose but vivid associations in the imagery and the fuzzy way in which we try to communicate through the systems do mirror a dream state to a degree,” says Dryhurst. “But I’m not sure dreaming quite captures it; it is perhaps more akin to derealization, the real sense that the world around you is less real.”

Artist and machine learning researcher Ryan Murdock, who makes strange pictures with his own custom-coded AI program, both is and isn’t in favor of the persistent analogies inspired by DeepDream. On the one hand, he echoes Dryhurst’s sentiment that the connection between machine learning systems and dreams is tenuous at best. On the other, he notes that thinking of AI through the language of dreams underscores its very real limitations. “The only thing I want a machine learning model to do is make art. I would never want one that uses a neural network at scale to make decisions about employment or anything important like that, even though many people try to apply them that way,” he says. “So I do think the language is really useful as a reminder: Don’t let a hallucinating dream machine make significant decisions!”

The False Mirror: AI’s Collective Conscious

Breton and his fellow Surrealists sought to channel the unconscious through their work; theirs was an art of ids, a pure portrait of the psyche. As a tool of mass application, AI art doesn’t lend itself to the same kind of internal excavation, but it does reveal something else. “For Surrealism, it was this idea of the collective unconscious—the dream state. But now, with all these scraped images and working with data, it’s more of a collective conscious,” says Carla Gannis, a New York–based artist and teacher who leads a class called “Subverting Digital Media” at New York University’s Tandon School of Engineering. Gannis has been working with AI in some form since 1998, but she is more excited about the creative potential of the medium today than she’s ever been. “There’s always this talk that AI is just a tool. But as a teacher, I know this is not like any other tool I’ve used for artmaking,” she says. “It’s learning from me.”

Over the past year, Gannis has been making automatic drawings by hand that she then feeds into AI platforms. “For DALL-E, I’ll type in something like ‘Please respond to this image with one of your generated images’ and upload a photograph of my drawing. We’ll go back and forth through several iterations.” What comes out in the end, she says, is both a distorted portrait of her own artistic impulses and a telescoped snapshot of the data the system used to recreate it.

“I’m very interested in the way [AI] makes pictures,” adds photographer Charlie Engman. He began casually experimenting with Midjourney a few years ago, and the psychic charge of the results quickly compelled him to dig in. “Essentially, it aggregates visual tropes from data and attempts to reproduce those tropes without any recourse to what we would call ‘experience.’ This means it mostly makes conventional or clichéd imagery, but it does so in strange and uncanny ways that reveal a lot about the expectations and value judgments we bring to the things we look at.”

Some of the most interesting AI art projects today involve either custom-made programs or the hacking of readily available ones. Most users want to either push the technology’s limits or draw winking attention to them. Engman is one of the rare artists to have used an AI image generator on its own terms to create works that feel authentically novel, images only he could make. His brand of Surrealism is the Magritte kind: minimal, wry, a mix of everyday scenes and otherworldly phenomena. One recent effort depicts the head of a woman emerging from the opening of a flower vase.

The Fuzzy Future

The long tail of Surrealism’s influence in art history owes as much to the movement itself as it does to its dissolution. The onset of World War II was the coup de grâce for the organized movement, but even before that, infighting and ideological disagreements among its members slowed the momentum, and many artists—Breton, Dalí, Max Ernst, André Masson—eventually dispersed from Europe to other parts. In the end, what made Surrealism so potent—its elusiveness, its malleability—is what made it difficult to contain and control, even for the people who invented it. But it needed to die to live anew. Once freed from egos and manifestos, Surrealism stopped being a movement and instead became an idea, a vibe, something artists could tap into and mold to the interests of their cultural moment.

Murdock sees this artistic potential in AI too, just as he sees the need to protect its accessibility from commercial interests. Right now, he explains, there is a wealth of open-source tools available that make bespoke projects like his possible. But as few and fewer companies gain more and more power in this field, those resources will continue to be closed off, bought up, or else made obsolete. “We continue to gain these capabilities,” he says, “but if we go completely proprietary with them, we end up not having the ability to do weird tinkering experiments with them. We cede that to these large corporations.”

“I believe creative expression is a shared endeavor that cannot be rightfully claimed by any one individual,” says Engman. “Creativity is a collective and incremental process. But we live in a world where the value of everything has been reduced to its capacity to produce capital. There’s so much potential in AI, but we have to work for it.”