Being reviewed:
Advertisement
What do photographs literally represent? And who decides? Such questions are often relegated to the ethereal realm of “cultural studies,” but—as the world now swims in images, used in everything from surveillance to LLM training to AI targeting of air strikes—the consequences are all too real. The missile strike on Shajareh Tayyebeh elementary school in Minab, southern Iran, raises a series of serious ethico-political questions; one of the most pressing and often neglected, however, concerns representation as such. The school in Minab was adjacent to, and previously part of, an Islamic Revolutionary Guard Corps (IRGC) naval installation—which was among a bundle of targets razed by the US in the early hours of the conflict. Although the school had been operational for years, its coordinates and outdated status remained encoded in the dataset as a military target: a deadly classification with dire results. Shajareh Tayyebeh elementary school appears to have become a target because of that encoding, itself informed by tools of object classification and recognition (what an IRGC naval installation “looks like” is what an IRGC naval installation is). This means that the potential sense of the site was determined and delimited in advance, dictated by satellite imaging. In this case, what the image represented—and to whom—was a matter of life and death.
Who, or what, “saw” these targets then, before they were struck? The answer is machines, not people, even though it was people who then bombed the facilities and people—children—who died. And this leads into the question infusing American artist Trevor Paglen’s How to See Like a Machine: What is an image if it is no longer even made to be viewed by human eyes, even as it acts on us?
The Belgian surrealist painter René Magritte is most well-known for a painting from 1929 titled The Treachery of Images. The painting is straightforward enough at the level of depicted content: a clear, literal rendition of a smoking pipe, along with a caption, “Ceci n’est pas une pipe” (This is not a pipe). In 1964, Magritte made a similar work depicting an apple. In his latest book, Paglen recounts how he fed an image of Ceci n’est pas une pomme to an AI system designed for object recognition and classification. Of course, the AI has no trouble declaring that, in fact, this is an apple.
Words and images alike are for Magritte porous entities that always fail to directly align with their purported referents “out there in the world.” They are malleable social conventions whose inability to stick—their inherent ambiguity—means there is always more to say or show in the fields of communication, expression, and even description. By contrast, Trevor Paglen’s “machine realism” is not concerned with any of these fields. Because classifiers are trained on sets of images that are already taxonomized, object recognition is always an extension of a preestablished, and artificial, harmony between image and referent, or sign and world. “Machine realism,” Paglen writes, “isn’t the result of recent technological developments but the digital reincarnation of an ancient impulse to reduce a complex and relational world down to measurable physical traits mapped onto rigid classification.” Through a process of accumulation, these technical systems construct the world in accordance with an essentialist logic steered by their prospective utility and marketability. An apple is an apple for consumption (and perhaps now, destruction) alone. This is why errors of classification can always be redressed for the sake of improving the overall integrity of the system, expanding its scope. Literalism “is a feature, not a bug.”
Photography (and film) inaugurated a paradigm shift in the field of representation, argued cultural critic and philosopher Walter Benjamin in his 1935 essay “The Work of Art in the Age of Mechanical Reproduction.” Photography didn’t just change science and aesthetics, argued Benjamin, but also changed politics. He was living through a cultural upheaval introduced by photography, but now powered by print technologies that allowed for images to circulate at rates hitherto unimaginable. Technological reproducibility meant rapid dissemination and freedom of access for both producers and consumers alike.
But what Benjamin was really highlighting was that this revolution did not inherently lead to democratization. The new age had also birthed more insidious forms of control—the proliferation of propaganda, an economy of attention, and thus new models for homogenization—that authorized a fascism that Benjamin was already seeing dominate.
Contemporary images rendered by computer vision and generative AI are no longer representational, they are activational.
Being inundated with images, even a century ago, led to an unaccountable, pervasive and totalizing spectacle, as well as ever increasingly surreptitious means of surveillance. All of this, of course, we are intimately familiar with today.
There is, however, a crucial difference between Benjamin’s vision and today’s visual culture. Contemporary images rendered by computer vision and generative AI are no longer representational, they are activational.
The difference is not between passive and active. Representation, of course, has a constructive component: what an image depicts is as significant as how the content is depicted. Visual representation has always been a site of contestation, precisely because it concerns meaning: how images become meaningful as they pass from creator to distributor to receiver (institutional presentation, aesthetic codes and strategies, regimes of significance). Representation is less a concept, then, than a complex order of dispute, a field of antagonisms that demand thoughtful engagement and, more often than not, resistance. Magritte’s genius was in bringing forward the insurmountability of this antagonistic truth, this treachery of the visual.
The crucial shift in recent visual culture is the deployment of images that are no longer meant to mean anything. Instead, they are meant to initiate direct action in the real world—or the world of the real, our direct environs—irrespective of their sense. The antagonisms of representation are superficially mended in activation, for the worse.
Machine vision no longer privileges either a human observer or a spectator. Images today are produced and consumed, first and foremost, by machines and for machines. Paglen writes:
A photograph shot with a phone creates a machine-readable file that does not reflect light in such a way as to be perceptible to a human eye. A secondary application, like a software-based photo viewer paired with a liquid crystal display and backlight, may create something that a human can look at, but the image only appears to human eyes temporarily before reverting back [sic] to its immaterial machine form when the phone is put away or the display is turned off.
This is the invisibility of contemporary images: machine-readable and hermeneutically mute at the same time. They are, in effect, nodes in an ever-expanding network designed to provide, and trade, information between systems concerned with the preferences (erotic, shopping, viewing, and the like), habits (travel, eating, and the like) and behaviors of human actors. Thus, images no longer exist for us. But every day, we live in accordance with a series of images we never see.
Such are the stakes of Paglen’s analysis of the state of “invisible visual culture” today. “The vast majority of the world’s images,” writes Paglen, “have become a means for technical systems to communicate with one another, remaining largely invisible to human eyes.” And this ubiquitous invisibility is the crux of the transformation from representational to activational that organizes the entire text: “In my thinking about the world of automated image-interpretation and image-making, I’ve found it useful to move away from the concept of representation and toward the concept of activation. Instead of asking what these technologies say about the world, I want to ask what they do.”
Paglen’s artistic output has been a staple of the art scene for the past decade and a half. His work stands at the forefront of critical engagement with the practical use of algorithms (Machine Readable Hito, 2017), the novel deployment of digital sensors for facial recognition and object classification (ImageNet Roulette, 2019), and these sensors’ histories as military applications (It Began as a Military Experiment, 2017). Now, in How to See Like a Machine, Paglen claims that since images are no longer functionally representational at the level of their technical operations, we can no longer rely upon our inherited means of image critique.
Most disciplines concerned with the social role of images (art history, media studies, film theory, etc.) have historically relied upon the theoretical armature of semiotics to explain and uncover how images take on their semantic value. When treated as signs, images as tools for the production and maintenance of ideology, as sites for affective investment and enjoyment, and as the means for recording and producing history are laid bare. But all these functions are representational: they generate a sense of the world for some subject viewing them or produce a view of the world for some subject making them. As meaning is inherently ambiguous, representational strategies always contain blind spots that can be exploited to complicate the reception of images, often challenging status quo assumptions and subverting expectations. For Paglen, however, invisible images do not mean anything, so there are no blind spots to exploit.
Returning to his AI experiment with Magritte, Paglen finds a paradox at the heart of this machine—or market—essentialism. Formerly, the relationship between perception and reality, for essentialism, gave the latter pride of place. A system of representation, in this view, whether physiological or aesthetic, maintains a mimetic relation to the world, not a constructive one. Twentieth-century investigations into the neurophysiology of vision, however, revealed the importance of emergent pattern detection for seeing. David Hubel and Torsten Wiesel won a Nobel Prize in 1981 for demonstrating that “individual neurons in the visual cortex act as feature detectors, responding selectively to precisely defined elements” such as edges, dots, and ellipses. The composite images that populate our visual field are often inferred hallucinations derived from these primitive elements.
Such is the foundation for the use of neural nets in machine vision. You can either take an image and break it down into its constitutive visual primitives and make a prediction about its identity or start with an image and have a neural net assemble an alternative by “generating synthetic primitives” and rendering them into novel forms. This is referred to as the “neural activation theory” of perception. The crucial point is that “in this conception, an ‘image’ is a name for the relationship between a pattern of excited neurons and our conscious experience of that activation,” but the activation itself is preconscious—or rather, unconscious, but not exactly in the psychoanalytical sense.
Curiously enough, Doritos are an instructive and central metaphor for Paglen. Everyone knows that Doritos are not really food; their nutritional value (if any) is entirely irrelevant compared to their stimulus value. Everything—from their color to their texture to their taste—is designed to “maximally stimulate our sensorium,” activating pleasure receptors in our brains with an intensity unmatched by anything “natural.” Their goal is “neurological engagement.” In this sense, they are activational. A feedback loop of consumption is the goal of a single Dorito, and therefore every Dorito. Neurological engagement, as the object of activation, prompts incessant hunger.
For Paglen, AI-generated images are akin to Doritos: their object of excitation is not the eye or even the mind, but the brain itself. Unlike all images in prior visual history, AI images are not meant to be beheld, contemplated, or scrutinized; they are meant to trigger automatic responses. In the same way that Doritos are engineered to induce more consumption, the machine images we consume are ordered toward maximizing our engagement with the technical systems that produce them. Bypassing conscious enjoyment and reconfiguring our neurological mechanisms of selection, Dorito-images are a psychological operation (psyop), for Paglen.
When light reflects off some object, passes through the camera lens, and inscribes itself onto a photosensitive surface (or sensor), a record is left of that object—and, crucially, of that object existing before the camera. This “indexicality,” a physical, existential relation of contiguity between a sign and its object, is the basis for treating photographs as evidence and documents.
In a mediascape of neural activation and invisible images, however, indexicality becomes nothing more than an aesthetic to be borrowed from, incorporated into the training data, leaving every possible image in a state of “superposition.” Undecidability becomes the modus operandi for every would-be viewer of any given machine-rendered image.
For this reason, the viewer experience of any given image now is equivalent to encountering UFO photos, precisely because they are sites where preconscious belief meets the visual codes of empirical observation. No one’s mind is ever changed about the reality of UFOs when given visual evidence, and yet, evidence is what the images are intended to signify as.
And herein lies the political crux of the activational regime: How can perception be sutured to belief in a society of unparalleled undecidability? Since “we are all UFO researchers now, trying to make sense of a world whose epistemic foundations have been swept out from under us,” by what means can we establish an ethics of seeing?
Images that move between machines are invisible to human perception, even while structuring our perception of the world in accordance with self-regulating protocols. Human perception, when facing the products of generative AI, is equally unable to visually detect the kind of rewiring wrought by these Dorito-images. In both cases, “the circulation of images is a largely automated circuit that reinforces itself through feedback mechanisms built into its systems.” Seeing is no longer simply a means to believing, but a “form of world making.”
This is the transformation in the treachery of images traced by Paglen. Belief in the literalism of a correspondence between image and referent used to be the target of critical work, both artistic and theoretical, to reveal the antagonisms of representation. The current architecture of machine vision, by contrast, no longer requires belief for this literalism to produce its deleterious effects. Seeing is no longer the privileged operation of a human viewer or primarily a matter of either critical or naïve reception. Seeing has been extracted and automated, and we are all now moving through a world no longer built for (and by) human eyes.
Even today, when vision is weaponized like never before, the architecture of machine vision remains invisible. Making it visible thus becomes the paramount task for encoding a new literacy in public consciousness, for recovering an ethics of looking. And this, in fact, is the same demand Benjamin made of photography’s critics at the beginning of the last century. It is to Paglen’s considerable credit that the text itself performs this task: the invisible is here legible, for anyone willing to look.
The question of what to do with that legibility, however, remains genuinely open. Paglen’s aim, in his text and his art, is to offer a new technical and critical lens. Even if semiotics is nearing obsolescence, Paglen’s new literacy is still more diagnostic than prescriptive. Even this diagnosis might be too late; present circumstances are reaching a point of critical mass.
Knowledge will not be the catalyst to change, unfortunately—even if it is a requisite ingredient. Instead of a strategy for combat, all we have is a map of the territory. And that may not be enough. 
