Generative AI menus have entered the restaurant business, but customers often sense something is off even when they can't articulate why. The models are trained on a narrow, “pleasing” aesthetic that produces illustrations that are either egregiously fake—like a burrito with cheese so bubbly it resembles avant-garde art—or so ordinary that something feels wrong upon closer inspection. Reality Defender CTO Alex Lisle described it as “almost like an alien trying to make a pizza without understanding its core principles.”
The problem stems partly from convergence, where AI models trained on similar sources reinforce each other's outputs. When AI generates a fast food menu, it references Wendy's, Burger King, or McDonald's—all of which share similar styles—creating a feedback loop that further homogenizes the results. This mirrors how Big Mac commercials arrange each layer by a prop designer to look maximally appetizing, but amplified in AI outputs.
When AI models train on too much AI-generated content, they risk model collapse. Lisle compared it to “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.” Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, explained that data sets are optimized for “pleasingness, or not being offensive, and so there's a way that turns into homogenization.” AI, he noted, “shaves off the edges” in both images and language.