Why AI-Generated Food Looks Inedible

Cafés, restaurants, and brands are increasingly commissioning promotional food images from generative neural networks — and the world is being swamped with appetizing-looking garbage. In The Verge's roundup: shrimp that resemble doughnuts, Reuben sandwiches that look like they came from the depths of the sea, worm-noodles, doughy pasta squiggles, and stringy chicken. There's also ice cream reminiscent of cracked concrete, and burgers that seem sculpted from stone. And everywhere — lumps, holes, endless holes.

Where the generation breaks down

Most leading generators work through diffusion: the model starts with a screen of pure noise and gradually removes it, reconstructing the picture. As Oxford professor Chris Russell explained, the coarse structures emerge first, and fine texture details are added at the very end. If an object's basic shape turns out wrong already at an early stage, vivid details get layered on top of it — the same glitch that produces a person with six fingers.

A separate headache is thin, elongated structures that trail off somewhere. According to behavioral scientist Giovanbattista Califano of Naples, noodles, threads, and tendrils are exactly the geometry that diffusion models stumble over: having started to draw something of the sort, the network doesn't understand where to stop or what to attach it to. It's just as bad at keeping repeating textures like bubbles and seeds within bounds — hence the obsessive “noodliness” and clusters of holes capable of triggering trypophobia.

The machine doesn't know what a sandwich is

The main reason runs deeper: the neural network doesn't understand what it's drawing. It has statistically absorbed how a sandwich, noodles, or a burrito usually look, but not why they look that way or how they behave in the real world. Zurich professor Roland Meyer notes that the generation “reproduces appearance without knowledge of the world” and perfectly imitates the look of a photograph, but not a photographer's professional techniques — and that is exactly what's troubling.

The training data makes the picture worse. Food photography is stylized: sharp contrasts, glossy light, exaggerated shapes, and sometimes what's shot isn't even food. On top of that, London's Simon Colton observes, no one uploads a boring red apple to the network, whereas strange, meme-worthy images spread far and wide. And training models on the output of other models leads to “collapse” — visual degradation and mounting uniformity.

Why it makes our skin crawl

The final touch is human perception. Scientists believe disgust evolved as a defense against parasites, pathogens, and toxins, which is why we're especially sensitive to wrong food. Wormlike tendrils recall parasites, clusters of holes suggest infection, unnatural colors signal spoilage. AI food hits all these triggers at once: it looks like food, but the brain knows better — and we instinctively recoil.

Source: The Verge