When it first happens to you, you think you’re crazy. You walk into a coffee shop and look at a menu with a variety of bagel sandwiches, but each illustration seems eerily flawless, precisely symmetrical and strangely smooth, provoking a visceral feeling that something is not right. You might think you’re paranoid, but you’re not losing your mind. Generative AI menus have come to the restaurant business thanks to models trained on a narrow, “nice” aesthetic that produces a look that feels off even when you can’t explain why.
Sometimes these illustrations are egregiously fake, like a burrito with cheese so bubbly and melty it looks more like avant-garde art than lunch. More often, they are so common looking that you only notice something is mistaken when you take a second to look closer.
“It’s almost like an alien trying to make a pizza without understanding its basic 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, a business that exists in part because of problems like this.)
Lisle says the way these models are constructed may help explain why the illustrations seem to take on such a specific aesthetic: one in which each scoop of ice cream is perfectly round and in which shrimp appear to have been genetically modified to eat their own tails, creating new “Lovecraftian food horrors.”
Large language models (LLMs) and diffusion models (the types of AI models that make chatbots and seemingly omniscient image generators like ChatGPT and Midjourney possible) are trained on large amounts of data. The models then identify patterns in the data sets to predict what a user is looking for when they ask something like, “Make me a menu for a burger restaurant.”
“A lot of these things look like the 2015 Chili’s menu, and there’s a reason for that,” Lisle said. “That was the corpus of work of which [the models] He drew his function.”
The new training data is invaluable to companies building AI models: Amazon has even been found sourcing rare books to scan and add to its training data, just to destroy those books once they have been uploaded. It is inevitable that some AI-generated content will leak into these incomprehensibly large data sets. But when AI models are trained with too much content generated by their own AI, they risk model collapse.
“Model collapse is almost like mad cow disease… when the results of a model feed back, eventually inbreeding becomes excessive and everything collapses,” Lisle explained. “What we see here is convergence, which is not necessarily a collapse of the model.”
Convergence is a little less extreme and degrades the quality of an AI’s results without making it completely useless.
If someone asks an AI model to generate a menu for a fast food restaurant, the model will likely reference menus from Wendy’s, Burger King, McDonald’s, or another popular chain. These menus already share a similar style, meaning that the AI-generated results will mimic that same style, only to reinforce it further if the AI-generated menu ends up back in the training data.
But menus and food advertisements will always look better than the real thing, like a Big Mac in a McDonald’s commercial where a prop designer arranges each layer of the sandwich to make it look as appetizing as possible. This effect may become even more pronounced in AI results.
“Optimization of data sets is to please, or you know, not to be offensive, so there is a way that becomes homogenization,” Lee Rainie, director of the Imagining the Digital Future Center at Elon University, told TechCrunch. “What AI is known to do in both images and language is remove edges.”
On a more localized scale, this image smoothing appears to occur when using an AI image generator to create a menu and apply edits to it. On X, a user named laboratorytec showed what happens when you create a menu in ChatGPT and then edit it 100 times to see the food looking less and less like it should. (We replicated the experiment and found similar results.)
“The end result really makes me uncomfortable,” Labtec wrote.
Restaurants are likely to fall victim to this problem and revise their AI-generated menus to alter small details over and over again, such as prices or item names. It seems like with each edit, the food images get a little rounder and softer.
“People have an almost inexplicable feeling when they look at something that’s generated by AI, compared to something that was real in the first place,” Rainie said. “There’s a sensitivity that people sometimes find difficult to express, 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] It’s so pronounced.”
There is science behind our aversion to these AI menus. Researchers at the University of Duisburg-Essen in Germany found that AI-generated food images exhibited an “uncanny valley” effectwhere images of food that looked almost real provoked more disgust and concern than images that were obviously fake. That apprehension only intensifies in light of the cultural context around AI.
If people react so negatively to these images, then that’s probably reason enough for restaurants to stop trying to make AI menus work. But the problems that perfectly browned hamburger buns bring us extend beyond the table.
“Seeing and hearing has always been believing, to the point that even our judicial systems are fully attuned to the idea that the gold standard of evidence is recorded confessions and video evidence,” Lisle said. “That is no longer the case. The world has fundamentally changed, for better or worse.”
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.