Why AI restaurant menus look weirdly perfect and slightly wrong

Shrimp curling into themselves, impossibly glossy buns, ice cream scoops so round they look engineered. There is a real technical reason AI food illustrations give you the creeps.

AI2Day Newsdesk4 min read
A glowing smartphone screen showing a grid of colourful Instagram-style photo thumbnails, surrounded by soft abstract digital light patterns suggesting data flo
Share

Key points

  • AI-generated restaurant menus are spreading fast, and diners report an instinctive sense of unease when they see them.
  • Researchers at the University of Duisburg-Essen found that near-realistic AI food images trigger more disgust than obviously fake ones, a well-documented psychological effect called the "uncanny valley."
  • Repeated editing of an AI-generated image pushes it further toward an over-smoothed, hyper-symmetrical look with each revision.
  • When AI models train too heavily on their own outputs, image quality degrades through a process researchers call "convergence."
  • The same forces that make a burger look surreally perfect online also have broader implications for trust in photos and video as evidence.

You walk into a cafe, pick up the menu, and something feels off. The bagel sandwiches look almost edible but not quite. Too smooth. Too symmetrical. Every sesame seed placed by a machine that has never been hungry.

You are not imagining it.

What is actually happening to these images?

AI image generators learn from enormous libraries of existing photos and illustrations, and most food photography is already styled to look better than reality. Think of a McDonald's commercial where a prop designer spends an hour stacking a single burger. The AI absorbs millions of images like that and treats them as normal.

The result is what Alex Lisle, chief technology officer at Reality Defender, a startup that builds tools to detect AI-generated content, described to TechCrunch as "an alien trying to make a pizza without understanding its core principles." The AI knows what ingredients should be present. It has no idea what it feels like to eat one.

The models these tools use, called diffusion models, meaning software that builds images from statistical patterns rather than actual perception, are trained largely on commercial food photos from chain restaurant menus. Lisle put it bluntly: "A lot of this stuff looks like a Chili's menu from 2015, and there's a reason for that. That was the corpus of work from which the models drew their function."

Why do the images keep getting stranger?

Editing makes it worse. Each time a restaurant asks the AI to tweak a menu, adjusting a price here or swapping an item name there, the images drift a little further from reality. A user on X demonstrated this by generating a menu in ChatGPT, then requesting 100 successive edits and watching the food become rounder, shinier, and progressively less appetising with every pass.

Lisle calls the underlying mechanism "convergence." When AI models feed too much of their own output back into their training data, the outputs narrow down toward a single heavily averaged aesthetic. Everything starts to look like the most statistically average version of "food." It is not quite the catastrophic "model collapse," where a system breaks entirely, but it produces what Lisle compared to a slow inbreeding effect, where variety and realism bleed away over successive generations.

Lee Rainie, director of the Imagining the Digital Future Center at Elon University, frames it differently: "What AI is known to do both in images and language is to shave off the edges." The optimisation is for pleasingness, not accuracy.

Should diners trust their instincts?

Yes. Human unease with these images is not just aesthetic squeamishness. Research from the University of Duisburg-Essen in Germany found that AI food images land in what psychologists call the "uncanny valley," a zone where something looks almost real but not quite, and that near-miss triggers more disgust than an obviously fake cartoon would.

For restaurant owners, the practical takeaway is straightforward: customers notice, even when they cannot say why. For everyone else, Lisle points to a wider concern. Photographs and video have long been the gold standard in courtrooms and newsrooms. AI-generated imagery is eroding that trust, one slightly-too-perfect shrimp at a time.

Common questions

Can you tell for certain whether a menu image is AI-generated?

Not always at a glance. Look for over-smooth textures, ingredients that curl impossibly into themselves, or a hyper-symmetrical arrangement that no kitchen could replicate. Specialist detection tools exist but are not foolproof.

Do restaurants know their menus look strange?

Many do not. They use an AI tool to create a menu cheaply and then edit it repeatedly, unaware that each revision nudges the images further from reality.

Is AI food photography always bad?

Not necessarily. A skilled designer using AI as one tool among several, with human oversight and real food photography as a reference, can produce convincing results. The problems arise when the AI works alone through many iterations.

© 2026 AI2Day