The first time it happens, you think you’re crazy. When you walk into a cafe and look at the menu of various bagel sandwiches, each illustration looks eerily perfect, precisely symmetrical, and strangely smooth, provoking an intuitive feeling that something is wrong. I may think I’m paranoid, but I’m not insane. Generative AI menus are hurting the restaurant business thanks to models trained on a narrow, “pleasant” aesthetic that creates a strange look, even if you can’t clearly explain why.
In some cases, these illustrations can be horribly fake. Like a burrito with cheese so bubbly and melty that it looks more like avant-garde art than lunch. Most of the time they look so normal that it’s only when you take a moment to look closely that you realize something is wrong.
“It’s like an alien trying to make pizza without understanding the core principles,” Reality Defender CTO Alex Lisle told TechCrunch. (Reality Defender itself is part of a growing category of startups selling AI detection and content verification tools; this business exists in part because of problems like this one.)
Lyle says the way these models are constructed helps explain why the illustrations seem to embrace such a particular aesthetic. All ice cream scoops are perfectly circular, and shrimp appear to have been genetically modified to eat their own tails, creating a new “Lovecraftian food horror.”
Large-scale language models (LLMs) and diffusion models (the kind of AI models that enable seemingly omniscient chatbots and image generators like ChatGPT and Midjourney) are trained on vast amounts of data. The model then identifies patterns in the dataset and predicts what users are looking for when they ask a question such as “Create a menu for a hamburger restaurant.”
“A lot of this menu is similar to the 2015 Chili’s menu, and there’s a reason for that,” Lyle said. “It was a collection of works that brought out (the models’) functions.”
For companies building AI models, new training data is invaluable. Amazon has even been found to source rare books, scan them and add them to their training data, and then destroy them after upload. It is inevitable that some AI-generated content will find its way into these incomprehensibly large datasets. However, if an AI model is trained on too much of its own AI-generated content, it risks collapsing.
“Model collapse is like mad cow disease…If you take the output from one model back into the model itself, you end up with too much inbreeding and the whole thing collapses,” Lyle explained. “What we’re seeing here is convergence, but it’s not necessarily a collapse of the model.”
Convergence is less extreme and reduces the quality of the AI’s output without rendering it completely useless.
If someone asked an AI model to generate a menu for a fast-food restaurant, the model might reference menus from Wendy’s, Burger King, McDonald’s, or other popular chains. These menus already share a similar style. This means that the output generated by the AI will mimic that same style, but only when the AI-generated menu is ultimately fed back into the training data will that style be further enhanced.
But food menus and advertisements always look better than the real thing. For example, in the Big Mac McDonald’s commercial, each layer of the sandwich is arranged by the prop designer to be as appetizing as possible. This effect can be even more pronounced in AI output.
“Optimizing data sets is not about being offensive, it’s not about being offensive, so there are ways to turn towards homogenization,” Lee Rainey, director of Elon University’s Center for Digital Futures Envisioning, told TechCrunch. “What AI is known for doing with both images and language is cutting edges.”
On a more local scale, this image smoothing seems to occur when creating a menu using an AI image generator and applying edits to it. In X, a user named Labtec showed what happens when you create a menu in ChatGPT, edit it 100 times, and watch the food keep changing in appearance. (I reproduced the experiment and got similar results.)
“The end result actually makes me uncomfortable,” Lovetek wrote.
Restaurants are likely victims of this problem, revising their AI-generated menus and changing small details like prices and product names over and over again. With each edit, the food image seems to become a little rounder and smoother.
“When humans look at something generated by AI, they get an almost inexplicable feeling compared to something that was real to begin with,” Rainey says. “Sometimes people have sensibilities that are hard to put into words, but you kind of understand it when you see it. I think that’s one of the reasons why some of the early backlash articles[against restaurants using AI menus]have been so prominent.”
There’s science behind our aversion to these AI menus. Researchers at the University of Duisburg-Essen in Germany have found that AI-generated images of food exhibit an “uncanny valley” effect. In other words, images of food that look almost real elicit feelings of disgust and anxiety more than images that are clearly fake. Considering the cultural context surrounding AI, this aversion only intensifies.
If people react very negatively to these images, that would be reason enough for restaurants to stop trying to make AI menus work. But the challenges of getting perfectly browned burger buns go beyond the table.
“Seeing and hearing is always believing, to the point where even our court system has fully adapted to the idea that the golden rule of evidence is taped confessions and videotaped evidence,” Lyle said. “That’s no longer the case. The world has fundamentally changed, for better or worse.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
