AI-generated menu images have quietly colonised the restaurant industry, and diners are noticing: not always consciously, but in a nagging, visceral way that something about the food on the page looks cosmetically wrong.
The mechanics are straightforward enough. Large language models and diffusion models, the technology behind chatbots and image generators like ChatGPT and Midjourney, are trained on enormous datasets. When a restaurant owner prompts one of these systems to produce a burger menu, the model draws on whatever burger menus, food advertisements, and stock imagery it has ingested.
‘A lot of this stuff looks like a Chili’s menu from 2015, and there’s a reason for that,’ Alex Lisle, Chief Technology Officer at Reality Defender, told TechCrunch. ‘That was the corpus of work from which [the models] drew their function.’
Why AI-Generated Menu Images All Look the Same
Lisle, appointed to the CTO role in July 2025, frames the core problem as convergence rather than the more catastrophic model collapse. ‘Model collapse is almost like a mad cow disease… when you feed the outputs from one model back into itself, eventually the inbreeding becomes too much, and the whole thing collapses,’ he explained. ‘What we see here is convergence, which isn’t necessarily model collapse.’
Convergence is subtler: the AI’s outputs degrade toward a shared aesthetic without becoming entirely useless. Every scoop of ice cream becomes a perfect sphere. Every shrimp curls in on itself with geometric precision. ‘It’s almost like an alien trying to make a pizza without understanding its core principles,’ Lisle said.
The training data itself pushes in this direction. Food advertising has always been aspirational: a Big Mac in a McDonald’s commercial is arranged by a prop designer to look maximally appetising, not like the item that lands in your bag. AI models trained on that corpus inherit the same idealising tendency, then amplify it.
Lee Rainie, Director of Elon University’s Imagining the Digital Future Centre, which was established in 2000 and expanded in scope in 2024, sees a systemic logic at work. Rainie spent 24 years directing Pew Research Center’s internet and technology research before joining Elon. ‘The optimization of the data sets is for pleasingness, or you know, not being offensive, and so there’s a way that turns into homogenization,’ he told TechCrunch. ‘What AI is known to do both in images and language is to shave off the edges.’
The effect compounds with every edit. A user on X demonstrated what happens when a ChatGPT-generated menu image is revised 100 times in succession: the food grows progressively rounder, smoother, and less anchored to any real-world kitchen. ‘The end result actually makes me uncomfortable,’ the user wrote. Restaurants iterating on AI-generated menus, tweaking prices or item names, are likely reproducing this drift without realising it.
Science Backs Up the Gut Reaction
There is now formal research to explain the discomfort. A study published in the journal Appetite, Volume 208, on 1 April 2025, by researchers at the University of Duisburg-Essen found that AI-generated food images produce an uncanny valley effect. Pleasantness, the researchers found, followed a quadratic function of realism: images that were either obviously fake or convincingly photographic fared better than those suspended in between.
Uncanniness, by contrast, followed a cubic function, rising sharply in that mid-range of near-but-not-quite realism. The full study, available in pre-publication form, involved a pilot with 99 AI-generated images varying in realism and a main experiment with 95 participants, averaging 31.28 years of age, who rated images on a spectrum from realistic to cartoonish alongside photographs of spoiled food.
The moderating factor was food neophobia, a wariness of unfamiliar foods, rather than food disgust sensitivity. The discomfort, in other words, appears to be driven by novelty aversion rather than any deep contamination-related instinct. Diners aren’t recoiling from AI menus because they think the food is rotten. They’re recoiling because something unplaceable is new and wrong.
‘People have an almost unexplainable sense about when they’re looking at something that’s AI-generated, compared with something that was real in the first place,’ Rainie said. ‘There’s just a sensibility that people sometimes find hard to articulate, but they kind of know it when they see it and I think that’s one of the reasons why some of the early stories about the backlash [against restaurants using AI menus] is so pronounced.’
Reality Defender, founded in 2021 out of a nonprofit research initiative and now carrying $33 million in Series A capital, counts IBM, Visa, and Comcast among its clients. The company uses a patented multi-model approach to detect AI-generated and manipulated content, a business category that exists in part because synthetic imagery is now spreading faster than human perception can reliably catch it.
Lisle’s parting observation stretches well beyond the restaurant industry. ‘Seeing and hearing has always been believing, to the point where even our court systems are entirely tuned to the idea that the gold standard in evidence is taped confessions and videotaped evidence,’ he said. ‘That’s no longer the case. The world has fundamentally shifted, for good or for ill.’
For restaurants, the more immediate question is whether diners who can’t name what bothers them about a menu will still vote with their feet. The uncanny valley research suggests the threshold is lower than most operators assume: near-realistic is worse than obviously artificial, and every iterative edit pushes the image further into the uncomfortable middle.
