Restaurants and brands chasing cheap visuals have flooded social feeds with AI-generated food that looks less like dinner and more like a biology class dissection. We’re talking donut shrimp, noodles that bleed into pastries, burgers that resemble masonry, and a burrito riddled with holes that could trigger trypophobia. The images are everywhere, and they’re almost uniformly revolting.
Chris Russell, a computer vision expert at Oxford, says the rot starts with how diffusion models actually work. These systems begin with pure static and strip away noise in stages — coarse shapes first, fine textures last. When the early structure is wrong, the model happily slathers vivid, appetizing-looking detail over a foundation that makes no sense. Russell compares it to those AI hands with six fingers: the broad strokes fail, but the skin texture renders flawlessly anyway.
Giovanbattista Califano, who studies AI imagery at the University of Naples Federico II, points to a specific geometry problem. Diffusion models are notoriously bad at generating thin, continuous, terminating structures — exactly what noodles, strands, and tendrils are. Once a model starts drawing something spaghetti-like, it can’t figure out where the strand should end or what it should attach to. The same weakness explains the repeating bubbles and seed patterns that spill across surfaces where they don’t belong, creating those clustered holes that make people’s skin crawl.
None of this should surprise anyone who’s watched AI image generators fumble hands, teeth, and text over the past two years. The models don’t know what a sandwich is — they only know the statistical appearance of sandwiches in training data. Roland Meyer at the University of Zurich puts it bluntly: AI reproduces looks without knowledge of the world. Michael Cook at King’s College London adds that training data likely includes highly stylized food photography with intense contrast and glossy lighting, which compounds the problem. The result is food that looks like it belongs in an architecture magazine, not on a menu. What’s genuinely baffling is that the actual food is presumably sitting right there in the restaurant kitchen, ready to photograph with a phone in under a minute.
