Why AI-generated restaurant menus all look uncannily the same

A recognisably wrong kind of appetising

Walk into a café lately and you may have noticed menu illustrations that are too flawless: bagel sandwiches rendered with perfect symmetry, cheese melted like something from an art installation rather than a kitchen. According to TechCrunch, this is not paranoia. Generative AI has arrived in the restaurant business, and its output carries a signature aesthetic that many people can sense but struggle to explain.

Reality Defender CTO Alex Lisle compared the effect to "an alien trying to make a pizza without understanding its core principles." His company sells AI-detection and content-verification tools, a market that exists partly because of problems like this one.

Trained on the corporate food photography of the past

The mechanics are straightforward. Diffusion models and large language models, the technology behind image generators and chatbots, learn statistical patterns from enormous datasets and then predict what a user wants from a prompt such as a request for a burger-restaurant menu. The aesthetic of the training corpus shows through in the results.

"A lot of this stuff looks like a Chili's menu from 2015," Lisle told TechCrunch, "and there's a reason for that." The models learned from the polished, chain-restaurant food photography that dominated their training data. TechCrunch also notes that appetite for fresh training material runs strong enough that Amazon has reportedly sourced rare books to scan, destroying them once uploaded, a sign of how far model builders will go for data, and of how easily AI output can seep into these datasets.

Lee Rainie, director of the Imagining the Digital Future Center at Elon University, frames the sameness as an optimisation problem. Datasets are tuned for pleasingness and inoffensiveness, he told TechCrunch, and that push "turns into homogenization." What AI tends to do, in images and language alike, is "shave off the edges."

Convergence rather than collapse

There is a related and more severe failure mode: model collapse, which occurs when models train heavily on their own AI-generated output. Lisle likens it to mad cow disease: feed a model's outputs back into itself and eventually the inbreeding becomes too much and the whole thing breaks down.

What menu imagery exhibits is milder. Convergence degrades output quality without rendering a model useless. Ask for a fast-food menu and the model will lean on the similar-looking menus of major chains; if that AI-generated menu later re-enters training data, the resemblance compounds.

Repeated editing compounds the damage

A widely shared experiment on X illustrates the mechanism. A user known as Labtec generated a restaurant menu in ChatGPT and then edited it 100 times, watching the food drift further from anything edible with each pass. "The end result actually makes me uncomfortable," Labtec wrote. TechCrunch says it replicated the experiment and got similar results.

That pattern likely mirrors what restaurants actually do: generate a menu, then revise it over and over to fix prices or item names, with each edit smoothing and rounding the food imagery a little more.

The uncanny valley of lunch

There is research behind the revulsion. Researchers at the University of Duisburg-Essen in Germany found that AI-generated food images produce an uncanny-valley effect: images of food that look almost real trigger more disgust and unease than images that are obviously fake. Rainie argues people have a reliable but hard-to-articulate instinct for AI-generated content, which helps explain why early backlash against restaurants using AI menus has been so pronounced.

Why it matters

The immediate lesson is commercial: if customers recoil at the imagery, AI menus are a liability rather than a cost saving. But the dynamics at play, convergence, homogenisation, and AI output feeding back into training data, reach well beyond the dinner table. As Lisle put it to TechCrunch, seeing and hearing have long been treated as believing, to the point that court systems consider taped confessions and video evidence the gold standard. That assumption, he argues, no longer holds. When the default aesthetic of generated content becomes uniform and detectably wrong, the larger casualty is trust in visual media itself.