Why AI-Generated Food Looks So Wrong (and So Revolting)

Key points

• Diffusion models, the technology behind most AI image generators, build pictures by removing random noise step by step, which often gets basic shapes wrong before finer details are added.
• AI image tools have no understanding of what food actually is, only a statistical sense of what it tends to look like in photos.
• Human disgust reflexes, which evolved to warn us away from contaminated food, make us especially sensitive to anything that looks even slightly wrong on a plate.
• Training AI on AI-generated images can cause quality to spiral downward over time, a pattern researchers call "model collapse".
• Businesses using AI food images risk putting customers off rather than drawing them in.

Something strange is happening on restaurant Instagram feeds. Shrimp with the proportions of donuts. Burgers that appear to be made from gravel. Ice cream that looks like cracked pavement. Noodles that stop making sense halfway across the plate.

More food businesses are turning to AI image generators, software that creates pictures from written descriptions, to produce promotional photos. A significant number of those images are going badly wrong, and food scientists, computer-vision researchers, and behavioural scientists now have a fairly clear idea of why.

How does AI actually make an image?

Most leading image generators use a process called diffusion. Start with a screen of pure visual static, then strip away the noise gradually until a picture forms. Chris Russell, a professor of AI, government, and policy at the University of Oxford, explains that broad shapes appear first and fine texture comes last. That ordering is where trouble starts.

The model might lock in a flawed overall structure early, then paste vivid, realistic-looking texture on top of it. The result is something that reads visually as detailed and food-like, but is structurally wrong in ways that are hard to ignore. Russell compares it to the well-known problem of AI-generated hands with six fingers.

Why are noodles and holes so particularly bad?

They trip up a specific weakness. Giovanbattista Califano, a behavioural scientist at the University of Naples Federico II who studies responses to AI imagery, puts it plainly: diffusion models struggle with thin, continuous lines that have a clear start and end point. Noodles, strands, and tendrils are exactly that geometry. So they bleed into the wrong places, attach to nothing, and multiply without reason.

Repeating textures, seeds, bubbles, pores, behave similarly. They spill across the image where they have no business being. That is why so many AI food images tip into trypophobia territory, a sensitivity to clusters of small holes that many people find deeply unpleasant.

Does the training data make things worse?

Yes, in several ways. AI models learn from images scraped from the internet, and that is an unusual diet. Food photography is already heavily stylised, full of extreme gloss, exaggerated colour, and sometimes fake food standing in for real dishes. Roland Meyer, a professor for digital cultures and arts at the University of Zurich, notes that AI picks up those surface aesthetics without understanding why professional photographers make those choices.

Beyond that, weird images spread further online than ordinary ones. A plain red apple gets no shares. A grotesque food meme gets thousands. So models can absorb skewed associations about what food "should" look like.

Michael Cook, a senior lecturer in computer science at King's College London, adds that AI systems are increasingly trained on other AI-generated material, a cycle that researchers have shown can cause "model collapse", where image quality degrades and outputs grow increasingly strange.

Should restaurants be worried about using these images?

Probably yes. Human disgust is a powerful and finely tuned system. Scientists believe it evolved specifically to protect us from contaminated or parasite-ridden food, which means we are acutely alert to anything on a plate that looks even slightly off. Califano argues the uncanny valley, the unsettling feeling of something that is almost but not quite right, is more visceral for food than for almost anything else, including not-quite-human faces.

Noodly tendrils read as worms. Clusters of holes suggest infestation. Wrong colours and textures signal rot. A promotional image that triggers any of those responses, even unconsciously, is doing the opposite of its job.

For now, the simplest fix costs nothing: photograph the actual food.

Common questions

Can restaurants just write better instructions to fix this?

Better prompts help at the margins, but the core problems are structural. Vague instructions like "make it look delicious" give the model little to work with, and even precise ones cannot compensate for a system that has no concept of what food physically is.

Will AI image generators improve enough to get food right?

They are improving, but the fundamental limitation, generating images from statistics rather than understanding, means errors will persist. Specialist fine-tuning on high-quality food photography has produced cleaner results for some tools, though artefacts remain common.