When it first happens to you, you think you’re crazy. You wander into a cafe and look at a menu with a variety of bagel sandwiches, but each illustration looks eerily flawless, precisely symmetrical, and oddly smooth, eliciting a visceral sensation that something isn’t right. You might think you’re paranoid, but you’re not losing your mind. Generative AI menus have hit the restaurant business courtesy of models courtesy of models trained on a narrow, “pleasing” aesthetic that produces a look that feels wrong even when you can’t articulate why.
Sometimes, these illustrations are egregiously fake, like a burrito with cheese so bubbly and melty that it looks more likeavant garde artthan lunch. More often, they’re soordinarylookingthat you only notice something iswrongwhen you take a second to look more closely.
“It’s almost like an alien trying to make a pizza without understanding its 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 — a business that exists in part because of issues like this one.)
Lisle says that the way these models are built can help explain why illustrations seem to embrace such a specific aesthetic — one where every ice cream scoop is perfectly round, and where shrimp seem to have been genetically modified to eat their own tails, creating new “Lovecraftian food horrors.”
Large language models (LLMs) and diffusion models — the kinds of AI models that make seemingly omniscient chatbots and image generators like ChatGPT and Midjourney possible — are trained on vast quantities of data. The models then identify patterns in the datasets to predict what a user is looking for when they ask something like, “Make me a menu for a burger restaurant.”
“A lot of this stuff looks like a Chili’s menu from 2015, and there’s a reason for that,” Lisle said. “That was the corpus of work from which [the models] drew their function.”
New training data is invaluable to the companies building AI models — Amazon has even been found to source rare books to scan and add to its training data, only todestroy those booksonce they’ve been uploaded. It’s inevitable that some AI-generated content will seep into these incomprehensibly large data sets. But when AI models train on too much of their own AI-generated content, they riskmodel 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,” Lisle explained. “What we see here is convergence, which isn’t necessarily model collapse.”
Convergence is a bit less extreme, degrading the quality of an AI’s outputs without making it entirely useless.
“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,” Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, told TechCrunch. “What AI is known to do both in images and language is to shave off the edges.”
On a more localized scale, this smoothing of images seems to happen when you use an AI image generator to create a menu and apply edits to it. On X, a user namedLabtecshowed what happens when you make a menu in ChatGPT, then edit it 100 times to see how the food continues to look less and less like it should. (We replicated the experiment and found similar results.)
“The end result actually makes me uncomfortable,” Labtec wrote.