The cheeseburger on the La Quinta Inn room service handout immediately made my AI antennae tingle. The image—oddly symmetrical, bright, and busy—had something artificial about it. Maybe it was the cheese dripping over the patty in suspiciously triangular form, or the unnatural placement of the accoutrements, but it looked as if a burger emoji made a wish to become real food, and that wish was only half-granted.
While off-putting, the faux burger wasn’t anywhere close to the gastronomic calamities that have been spreading online lately. A buffet of viral food images have been circulating (again) and none of them are appetizing. What goes viral is the worst of the uncanny worst: incomprehensible chicken thighs, clustered hive-like burritos, bizarrely circular shrimp.
While the images are the opposite of appetizing, there’s a comfort in our ability to immediately clock them as AI. But experts say that’s quickly changing. AI has already permeated much of the food styling and photography world (at least outside of high-end production), and it’s only going to become more common and more sophisticated.
As AI image generators sort out their kinks when it comes to rendering food, we’re on the cusp of a system completely dominated by AI.
Why AI image generators struggle with food
Imagine a typical bowl of penne pasta: Tubular noodles lie in a helter-skelter pile, covered unevenly in sauce and grated Parmesan cheese. As a piece of imagery, pasta is textured, random, and full of varied details—and it’s exactly the kind of thing that AI image generators have historically had trouble reproducing.
According to Ahmed Elgammal, professor of computer science at Rutgers University-New Brunswick, current off-the-shelf image generators tend to use a stable diffusion model, which compresses images to save computing power.
“You are actually generating a very, very small tiny image and then you’re blowing it up later and trying to fill in the details,” Elgammal says.
Treating those missing features like an afterthought leads to strange results. “That’s where food basically falls apart,” Elgammal says. “Food is full of details, and our eyes are looking at all these details because that’s how we tell good food from bad food or how we tell the quality of the food.”
Many of the surrealist AI food images we’re seeing are the result of a circular system of unrealistic food imagery that has been around for decades.
Long before AI, companies would use practical effects. Campbell’s Soup famously added marbles to its bowls in advertisements to ensure its ingredients rose to the surface—one of a cornucopia of tricks food stylists have up their sleeves to make a photo drool-worthy through less-than-edible means.
And the addition of Photoshop and other postproduction digital technology made it even easier to render a perfect-looking meal.
The images conjured by AI draw on that history of unrealistic depictions of food. To come up with a yolky egg sandwich or a tuna melt, for example, AI models pull from preexisting images of these dishes and similar food items. Many are hyper-stylized, so they already have a sheen of the less real; others are preexisting AI images with their own issues. When they’re averaged together, they can create something a bit too otherworldly to pass muster.
Food has “so many variations,” says Hanseok Ko, an engineering professor at Catholic University who focuses on AI. “Different textures, moisture, different translucency, fiber, things like that, or even fat distribution—all of these things, in fact, introduce different forms and different images, and to be able to capture that by using just images as training data is almost impossible.”
Food images are already awash in AI
The images that go viral get our attention for one key reason: They’re downright nasty. But there are more stealthy ways that AI has already become a key part of food image production. A National Restaurant Association report found that more than one in four restaurant operators use AI tools at their eateries, with marketing as the most common application.
Small businesses and local restaurants are bearing the brunt of the AI food slop blowback, primarily because they’re using off-the-shelf technology that requires time, effort, and intentional prompting to produce something that resembles edible food.
Many large food service companies include AI in their workflow, too, but that rarely gets the same attention “because the process is way more intricate than what a lot of the mom-and-pop shops are doing,” says Rory Flynn, founder of SystematicAI and host of the Fast Hours podcast. He notes that big businesses are investing in custom-trained models that integrate their brand guidelines and draw from their previous food photography to keep their images looking consistent.
Even when companies spend the money on an actual photo shoot, AI still works its way into the process, especially in pre- and postproduction. Kristina Wolter, commercial food stylist and founder of GirlGoneGrits, has already felt the shift in her line of work.
While preproduction meetings with customers used to include stock images to get a sense of what the result ought to look like, they now feature AI-generated pictures. “That’s not a pot roast, that’s a rock inside that pot,” Wolter says. “I notice those things right off the bat or I’ll say the scale is always off.”
Wolter shoots far more “solos,” or images of food without a background, than she used to because clients will add their own mise-en-scène using AI later. “The company’s figuring out how to utilize taking those solos and putting them in any kind of backdrop that they want,” she says.
Wolter says she’s torn. She wants to use technology to help her do her job, but feels like AI is a threat to her business. “It’s a real challenge and a real fine line going into this world of AI because you’re trying to figure out, well, how do we manipulate AI so it actually does work to our advantage but not take the job completely away from us?” she says.
She acknowledges that this is far from food advertising’s first brush with manipulating image-making. Wolter wheels a huge tool kit to her shoots, filled with items like glycerin to create soda bubbles and glue to mimic milk for cereal.
“We still use some of those fake things to get that story across. But as a whole, we’re telling this beautiful story in this image,” Wolter says. “That’s an enhancement, but it is not the faking of the product.”
What comes next
Experts say we’re likely to see an explosion of more realistic food imagery. The reason is two-fold. The first is physical AI, says Ko, referring to the method of including physical properties among 2D images and text to train AI models. Count Nvidia CEO Jensen Huang among the most prominent proponents. The idea is that the “output becomes more meaningful, more realistic,” Ko says. Suddenly those issues with textures and lighting would become less of an obstacle.
The second is moving from a stable diffusion model to pixel diffusion. Under the latter, the model generates the image at the pixel level, meaning it can build in the tiny details from the outset rather than add them in later. These models are much more expensive, Elgammal says, because they require more memory and resources. But advances in hardware will make the tech more widespread. “I imagine these models will become more available in the next couple of years,” he says.
It’s not particularly appealing to look at grotesque images of food. But there’s something strangely honest about it. We’re genuinely debating whether AI might hasten human extinction, so seeing the technology fail so epically at rendering a realistic-looking shrimp is reassuring.
Food slop is a reminder that there are humans behind these questionable decisions. It’s an admission that time and money are finite resources, especially for small-business owners, and that technology is not infallible—not yet, anyway. So maybe, in our own perverse way, we should cherish the harrowing images of this moment. Because in the not-too-distant future, an entirely AI-generated shrimp will probably look like an actual shrimp.