What if The Atlantic owned a train car? I wondered. Amtrak, I had just learned on the internet, allows owners of private railcars to lash onto runs along the Northeast Corridor, among other routes. “We should have a train car,” I slacked an editor. Moments later, it appeared on my screen, bright red with our magazine’s logo emblazoned in white, just like I’d ordered. It’s an old logo, and misspelled, but the effect was the same: A momentary notion—one unworthy of relating to someone in private, let alone executing—had been realized, thanks to DALL-E 3, an artificial-intelligence image generator now built into Microsoft Bing’s Image Creator website.
This is now a habit. My colleagues and I have been asking Bing to produce Atlantic magazine covers on ridiculous topics, such as “The Burger Doctrine” and “What Cheeseburgers Want,” absurdist riffs on our political reporting. The AI has obliged (but couldn’t spell doctrine). Another of our generated cover stories, “The Case for the Cincinnati Bengals,” earned a luxurious teal cover with a rendition of the White House flanked by patterns of orange tigers in Art Deco symmetry.
In recent days, I’ve ushered any and every notion my brain produces through AI. Hearing a brief mention of the strange animal-rotation CAPTCHA used by LinkedIn and other sites led me to produce a picture of a hypothetical card game based on it. A social-media flare-up over New Yorkers’ disgust at the prospect of scooped-out bagels inspired a black-and-white photograph of a 1930s Manhattan worksite bagel excavation. My daughter texted, asking what her “goth name” should be; moments later, I sent back an Edward Gorey–style illustration of her possible Victorian-dead-girl alter ego. One colleague began texting me various items enclosed inside of chocolates—the Eye of Sauron, more train cars, Atlantic magazines, cheeseburgers. Other friends send new findings: instruments made of baby back ribs, an oil painting of Jesus Christ eating nachos.
It’s been possible to generate images from textual prompts with machine-learning-trained tools for some time now. When I first began reporting on this strange new technology, back in 2019, I described its early output as “just another style, bound by trends and accidents to a moment that will pass like any other.” Earlier this year, I said that more recent iterations were stupid—in a good way. But improvements to the tools have radically improved the quality of their output, and they have become easier to access too. The results have completely changed my view on what AI image creation means. It’s not for making pictures to use, even if that might happen from time to time. Instead, AI images allow people to visualize a concept or an idea—any concept or idea—in a way previously unimaginable.
[Read: The AI-art gold rush is here]
Since the generative-AI tide rose last year, worries about its uses and abuses have surfed its waves. A popular matter for debate: Could AI put artists out of work? Just wait. You won’t need photographers or illustrators anymore, some surmised. Art is a human practice that will always demand a person’s agency, retorted others. And in the murk between, a moderate position emerged: AI will change—not end—art practice, just as pigment, photography, and software had done before.
Each of these arguments relies on an assumption that now seems shaky: that the images AI generates would be used in contexts where images already find use. At the top of articles or printed inside magazines like this one, perhaps. As advertisements on billboards or on Instagram. Maybe inside mailers or on corporate websites. Perhaps even as fine art in galleries, or as prints on Etsy. All of that is already happening to some extent; AI pictures are finding their way into stock photography, glossy magazine spreads, and state-fair art shows. But to imagine AI as a mere outsourcing tool for picture work understates the weirdness of the technology.
For a time, auto-generated images were too broken to take seriously: people with seven fingers and ghoulish faces, the gestures of inscription in place of real, legible text. Many of those defects have already been overcome—the finger problem is more or less solved, and text is now possible, even if not reliable (“Rotote the Animal,” reads the box for my generated tabletop CAPTCHA game). Further improvements in accuracy seem inevitable, and still, the utility of generated images feels low—especially for those of us who don’t often require pictures for our vocations, or even our avocations. My colleague asked Bing for a Sanrio-style caricature of Ian Bogost, and the result looks just like me—I turned it into my iPhone contact photo and Slack profile pic. But that’s a pretty modest use case; it’s not as if I’d have commissioned such an image under different circumstances.
[Read: Generative art is stupid]
Mostly, I’ve been using Bing Image Creator just to see what any idea my brain conjures might look like were it given material form. I ask for these results to see what they would look like if they were real: an album cover for Burger King (Taylor’s Version). A professional awaiting the subway in 1960, wearing an Art Nouveau–ironwork men’s suit.
In previous eras, I might have turned concepts such as these into images via Photoshop. Ten years ago, I’d spend an hour ginning up a record cover for an album called Foucault Me Maybe, in which the French philosopher Michel Foucault replaced Carly Rae Jepsen, or designing a fake photo from Skeletor’s presentation at TED Eternia, superimposed with an inspirational quote from his talk (“The universe is power, pure unstoppable power—and I am that force; I am that power”). These are jokes, of course, but jokes are also serious. They invite people to see their world differently, at least for a moment. In so doing, the structures we take for granted are revealed to be accidental.
[Read: AI has a hotness problem]
Perhaps it isn’t right to paint my use of AI image generators, mostly to put cheeseburgers in or on things, as a serious intellectual pursuit. But I’ve certainly noticed that the technology works best when I use it to extend my imagination rather than my image generation. Seeing an Atlantic cover about cheeseburgers helps construct an idea of what such a topic might look like if taken seriously enough to be granted that pride of place. In so doing, it orients me to our publication’s purpose, even if I have no intention of producing a feature about the politics of burgercraft. To construct a goth rendition of my daughter demands pondering the nature of her personality and how it might be expressed in dark-horror form. Seeing the results teaches me something about myself and my experience, even if I’m the only person who ever sees the results.
Using AI to create real outputs—copies or amplifications of actual objects, scenes, or events—feels harder than allowing it to amplify my imagination. Here’s an example: A local bar near work recently started putting out bowls of some unholy snack mix. Seasoned nuts are in there, sure, but also yogurt raisins, dried banana chips, loose Mike and Ike candy, the occasional solitary whole corn chip, fully wrapped Jolly Ranchers. Imagine if a jovial giant ran through a snack-foods factory and then emptied his cuffs onto your happy hour. If you said an AI generated this snack bowl, nobody would bat an eyelash, my friends and I quipped over beers.
But when I tried to make good on the joke, generating such an image with Bing, I couldn’t provoke the AI to success. Its results were competent, but they were too organized to match the chaos of the original—the yogurt raisins were grouped together on one side, the whole bowl seemingly produced by choreography rather than bedlam.
I won’t presume to opine on the best use of generative-AI images, nor would I be so foolish as to try to predict their future. But the bowl of snack mix made me realize that approaching generative image creators in order to produce a desired result might get their potential exactly backwards. Instead, try spilling your unfiltered thoughts into its engine. AI can give them shape outside your mind, quickly and at little cost: any notion whatsoever, output visually in seconds.
The results are not images to be used as media, but ideas recorded in a picture. For myself, not for others—like the contents of a notebook or a dream journal. Thinking back, AI images have always served this purpose. Back in the early days of Midjourney and DALL-E, I remember seeing lurid pictures of (fictional) Art Deco automobiles, Streamline Moderne domestic kitchens, furnishings shaped like fruits. Beautiful, I thought, but what were they for? But that wasn’t the point, even then—and it certainly isn’t now. These aren’t images; they’re imagination. Imagine a leather-and-chrome espresso machine. Imagine a watermelon armchair. An impossibility, sure, but once thought—once visualized—an object for consideration and whatever that entails.