Why AI Food Photography Is So Unappetizing

The other day I found myself, as one is wont to do, wasting 10 minutes by playing around with an AI image generator. I was hungry at the time, and so eventually I began creating options for a hypothetical lunch: a shadowy charcuterie platter, rising up like the ruins of an ancient city with a sunset in the background; rings of rigid-looking calamari, seemingly made from lucite or glass, arranged in an artfully askew stack; and a circle of 12.5 cartoonish, smooth, translucent-red shrimp under a banner of cursive text that read, simply, “Shimp.” Some of the images looked like food; none of them looked edible.

As my lunchtime experiment showed, getting AI to generate a quality image requires knowing what you’re doing — starting with well-written prompts (beyond just “plate of shrimp”), a crucial step I had not taken. Sometimes, the results are amazing, like the AI-generated images Bon Appétit recently commissioned from artist Bobby Doherty, which accompanied a piece about an editor’s conversation with ChatGPT as it developed dishes for a hypothetical New American restaurant. Some of AI’s ideas for the menu were eye roll-inducing, as can be the case with New American restaurants, but Doherty’s vivid, otherworldly art still looks good enough to eat.

It would seem, however, that the average AI-generated food image is not quite there. In various corners of Reddit and Google Images, pizza slices and leaves overlap strangely or blend into each other, curries shimmer around the edges, turkeys have unusual legs in unusual places, and other supposed foods aren’t identifiable at all. On Adobe Stock, users may monetize AI-generated art, provided they have the rights to do so, and label their uploads as illustrations. Most of the platform’s photorealistic still lifes and tablescapes are passable, though a few veer into the grotesque: an endless ring of shrimp, all body and no head, or its impossible cousin with heads on either end. Images like these, and even ones that are less absurd, often reside somewhere in the uncanny valley — a much-debated locale that looms large in many conversations around AI.

Still, as tech companies tout AI’s applications for recipe development and even teaching cooking techniques, artificial neural networks are also making their entrance into the world of food photography. Some stock photo agencies, including Shutterstock, have partnered with AI platforms on their own image generation tools. Startups like Swipeby and Lunchbox intend to court restaurants and delivery operations in need of visuals for their online menus. Of course, a way to create visuals — paying food photographers to do their jobs — already exists. And beyond that ethical morass is a more immediate legal problem: Some AI models have been trained with creative works, often unlicensed, scraped from the internet, and can respond to requests to mimic specific artists. Understandably, the artists are starting to take things to court.

All moral concerns aside, for the time being, at least, food still looks most reliably delicious in the hands of food photographers, videographers, and food and prop stylists. So what is AI getting wrong? Karl F. MacDorman, a scholar of human-machine interaction and associate dean at Indiana University’s Luddy School of Informatics, Computing, and Engineering, says there are many theories as to what might cause certain representations to elicit feelings of eeriness or unease as they near full accuracy. “The uncanny valley is often associated with things that are liminal,” MacDorman says, as when we are not sure if something is alive or dead, animal or non-animal, real or computer-animated. This can be especially pronounced when an image mixes disparate categories, or assigns features to a subject that usually belong to very different things. It’s perhaps unsurprising that AI, at this relatively early juncture, might struggle with all of this.

While the original uncanny valley hypothesis, posited in 1970 by roboticist Masahiro Mori, was concerned only with humanoid figures, other uncanny valleys have since been demonstrated. There can be a similar effect with renderings of animals, and in a 2021 study, MacDorman and psychologist Alexander Diel found that houses can be uncanny, too. MacDorman suggests that food, likewise, has the capacity to be uncanny because of how intimately it’s connected with our lives.

John S. Allen, author of The Omnivorous Mind (published in 2012), has explored that connection from both a scientific and cultural perspective. An anthropologist who specializes in the evolution of human cognition and behavior, Allen speculated as to why some AI food can be so off-putting. “The familiar but slightly-off images are maybe the most disturbing,” he wrote in an email after I sent him some of my weirdest finds. “Maybe I interpret those in the same way I might look at something that I would usually eat, but which has spoiled or become moldy or is harboring a parasite or is in some other way not quite right.”

In The Omnivorous Mind, Allen argues that young children develop what he deems a theory of food (“sort of like a first language,” he says) that is shaped over time by varied experiences and cultural influences. “Our first visual impressions of what we eat set up expectations, based on experience and memory, about what something should taste like or whether we will like it or not,” Allen says. “When the food looks off, that sets up a negative expectation.”

MacDorman’s research supports a similar idea. When it comes to “configural processing” — simultaneously responding to many features at once, as with face perception — he says humans do rely on models we’ve developed of the food that we’re eating. “We have a model of what a shrimp should look like, what’s a good example or a bad example of shrimp,” he explains. If you see a shrimp that’s strangely long and thin, it’s not uncanny because it’s novel; it’s uncanny because it’s bringing to mind a familiar model, and when we try to fit them together, “there’s something definitely not meeting your expectations.”

Still, MacDorman thinks there can be feelings other than uncanniness at play in an adverse reaction to an AI-generated food image. “It could even be empathy,” he suggested. With a headless shrimp, for example, “you might feel bad because you wouldn’t want to be it.”

Some foods may provoke stronger reactions than others. “For me it’s the meat, all the way,” says San Francisco-based food photographer Nicola Parisi. “I do think meat in general is a very hard thing to photograph, even as a human, and I can see some of the same struggles with AI.” She thinks it also has yet to master other things some humans have trouble grasping, like composition, styling, and staying on trend. A dated backdrop, or a plating technique that’s no longer en vogue might not trigger any deep psychological phenomena, but they can certainly contribute to an overall value judgment of an AI-generated image. “A photo can be taken with a nice camera, and you can light it well, but it can be boring, or the styling won’t be great,” Parisi says. “A high-quality image could still be bad, you know what I mean?”

Luckily, there are professionals out there who know how to make food look great every time, and unlike AI, they can actually eat.

Hannah Walhout is a writer and editor based in Brooklyn.