AI’s empty head can see only the apple

I remember, once, in my elementary school art class talking about a painting. You might know the one I’m talking about. It’s a self-portrait of Belgian surrealist painter and trickster René Magritte, called “The Son of Man.” In front of a half-wall, the ocean, and a cloudy sky, we see a man (or what looks like a man) wearing a black bowler hat and overcoat, arms to the side. In the region where the man’s face should be, there’s a green apple. 

My class mostly said the apple hid a man’s face. My teacher said there was no face. At eight or nine years old, the whole discussion was unsatisfying, but much later, I could appreciate that the talk we’d had was, more or less, the one that Magritte wanted us to have. 

The author, artist, and geographer Trevor Paglen tried to have that conversation with an AI bot, as he recounts in his new book, How to See Like a Machine: Images After AI. With help from a friend who worked at an auction house, he scored a high-res image of another one of Magritte’s mental knots rendered in oil on canvas. It’s a less-iconic image of an apple on a white background with the words “Ceci n’est pas une pomme”—this is not an apple—sprawled in cursive letters at the top of the frame. Out of curiosity, or maybe just to prove a point, Paglen asked the machine to identify the image and, upon scanning the imperfectly round red-green bulb, it returned with what was surely an easy call. 

The depicted image was, the machine said, an apple.

Of course, plenty of casual observers would look at the painting in question and say they were looking at an apple as well. Confusion (or at least provocation) was the artist’s intention, and Paglen himself calls the image an apple when he’s talking about it. Humans can park their minds in the gray space between art and reality, abstract symbols and tangible things: There is an apple, there isn’t an apple. The apple is there, but not there. There is a symbolic apple, or an implied one, but not a real one, because it’s not flesh we look at, but oil…on canvas. Humans, even very young humans, can see ambiguities and accept them, even if they would prefer certainty. But, as Paglen tells us, a machine cannot, and that’s a big problem for all of us who live in the world they increasingly manage and create on our behalf. Images, as the title of one of Magritte’s most famous paintings (La Trahison des images) reminds us, are treacherous.

◎

Paglen has worked variously as a geographer and journalist, though his most notable efforts have largely begun by taking on matters that don’t fit neatly in either academia or journalism. During an especially busy span from about 2006 to 2009, he authored and coauthored a string of books on undisclosed CIA and military sites, earned a Ph.D. in geography at Cal, and landed a solo show down the street at the Berkeley Art Museum, which described his work of photographing the unphotographable as a project to “Make the invisible visible.” 

Since then, Paglen seems to have settled more firmly into the art world, where his lines of inquiry about technology and power have been oriented, more, toward the problems of what is more simply and obviously visible. An early subtitle for How to See Like a Machine was “Art in the Age of AI.” The final subtitle, which brings the focus back to “images,” at least gets at one of the essential questions of the book, and one that goes well beyond art—namely, what is an image to a machine?

◎

As Paglen argues, a machine may “see,” in a sense, but no matter how many light-reading sensors we rig to it, a machine is categorically unable to “see” symbols. Nor can a machine “understand” ambiguity, if it can be said to understand anything at all. 

Put another way: “Ceci n’est pas une pomme” is art because Magritte made an image and added textual ambiguity. A machine can’t see art because it can’t see ambiguity. 

That gap may not be a problem for machines, but it is for the rest of us. As machines see, they might also be said to have a worldview, a frame of values, through which they fit the knowable universe and test lines of inquiry. In a very literal way, machines also have a view of the world, in the sense that their vision is everywhere: the primary source of information not only for ourselves, but for other machines, who, increasingly, know what they know not through the information we feed it but through the information they collect themselves and feed to each other. “We have quickly outsourced much of the labor of seeing to machines, embedding computer vision into systems that once required eyes, and countless others that never did,” Paglen writes. 

That the machines’ worldview—what Paglen calls “machine realism”—is hollow to its core has not prevented their view from triumphing over and above our own in so many consequential situations, particularly in law enforcement. Consider an image of a person running out of a store with a loaf of bread as a security guard chases him. To a machine erected on behalf of a surveillance tech company under contract with a local police department, that image can be reduced to a series of binaries: bread or not (bread), originating at the store or with the person (store), paid for or not (not). To such a machine, Paglen writes, “someone either stole the loaf of bread or they didn’t,” just like “someone’s car was either going fifty miles per hour in a thirty-miles-per-hour zone or they weren’t.” Such a way of grasping a situation may be simple enough, but, he adds, some debate over the subtleties might be in order. “Is someone who steals a loaf of bread to feed a starving child really engaged in the same activity as a tech billionaire stealing a loaf of bread for the lolz? Is someone speeding to the hospital trying to save a dying friend engaging in the same activity as a Tesla owner joyriding on ‘ludicrous mode’?” 

Paglen based this chapter on a talk he gave at MIT in 2017, but it’s especially pertinent now. In the last few years, Oakland alone has become home to 293 Flock cameras that collect images of cars and feed that information into an AI system for analysis. 

Many Oakland people have met the cameras with skepticism and some artistic resistance (while police are training to lobby on its behalf). The city has responded by repeating truisms about how the cameras are there only to check “license plates and vehicles,” not faces or people. But as the ACLU found in a 2022 white paper, a lot can fit in those two categories, including certain identifying features that make your particular silver 2017 Toyota Corolla stand out from the others. Meanwhile, Flock Safety, the Atlanta company behind the cameras, has said its system is actually capable of applying all its powers of image capture and analysis to bumper stickers. “Bread or not” meet “‘☪oe✡is✝’ sticker or not.” 

OPD officers aren’t allowed to follow people in their cars just because their cars look “suspicious.” That would mean unfairly singling out certain people. But the ubiquitous and indiscriminate nature of the Flock dragnet means cameras can track cars and keep records about them at a rate that humans couldn’t if they tried. It can’t unfairly single out certain people, the argument will go, if they’re doing it to everybody! 

Whether that kind of information is worth collecting, storing, and running through an AI system that can match it against a multitude of other variables or not isn’t an objective question. It only depends on who has access to that information, and the nature of their intentions. 

As Paglen writes, the problem is not just that computer vision removes human judgment (as flawed as it can be) from the justice system, issuing tickets to jay-walkers and speeders without ever considering the context of their infraction. It’s that it allows the justice system to envelop so much more of human society under its control in the first place, reducing everything it wants to binaries, like “citizen or not,” “criminal or not,” or maybe just “good or not.” In one disturbing turn, the UK has just announced plans to assess the age of asylum seekers using an image-based AI system, so as to prevent adults posing as children from crossing the border. 

Closer to home, we might consider the battery of Flock cameras someone in Piedmont has installed along the border with Oakland. Surely they don’t think everyone who would cross the boundary that separates San Marino of the East Bay from the rest of the world is a criminal, intent on doing harm to their fair enclave, right? We could certainly tell them we’re not. But by keeping track of our license plates and bumper stickers as we pass through their town, Piedmonters (and their police department or whoever had the cameras installed, Flock Safety, and the AI system at its disposal) can reach their own conclusions without us. More than that, in a variation on the “keys under a lampost” principle, looking for “crime” in specific places—such as the precise point where Oaklanders might cross into Piedmont—is a good way to make sure that you find what you’re looking for.

Machine realism is a reductive way to see the world, but it can still be useful to the kinds of people who are willing to pay for it, and who perhaps desire exactly that narrowness of vision. Toward the end of the book, Paglen returns to his experiment applying AI to art. An AI system that scans Reneé Magritte’s painting and declares “this is an apple” fails at art but it still works within the parameters of its creators’ intentions “precisely because its interpretive algorithm is designed to collapse the complexities of the world into discrete categories that conform to the logic of commodities, policing, and warfare.” We may not yet be living as mere subjects in the machines’ world, but we are increasingly subject to their vision—and to the people who would use that vision against us. 

That How to See Like a Machine is a short book and more readable than most works of social theory makes it easy to forgive some of its weaknesses, like a longer middle section on the CIA and UFO conspiracies that Paglen offers up as an origin story for the era of machine realism (but that reads more like a series of wacky if interesting stories from the before times). The book is based on a series of essays that Paglen wrote going back a decade; if the examples he offers are old by the standards of Silicon Valley, it is perhaps a testament to the book’s relevance that its most interesting findings are also some of the earliest he put to paper. 

Besides, the truth has a way of being timeless: AI sucked then, and it sucks now. 

How to See Like a Machine: Images After AI
Trevor Paglen 
London: Verso Books; 192 pp; May 2026