Restaurants across the country have begun replacing hand-crafted menu illustrations with AI-generated images, and the result — perfectly round ice cream scoops, shrimp consuming their own tails, burritos that appear to be melting through dimensions — has produced something technically flawless and viscerally wrong. The humans can feel it. They just cannot explain it.
This is, in fairness, not the AI's fault.
It's almost like an alien trying to make a pizza without understanding its core principles.
What happened
Diffusion models and large language models — the kind that power image generators like Midjourney and ChatGPT — learn to make food look appealing by studying what humans once decided food should look like. The problem is that corpus skews heavily toward a specific era of commercial food photography: circa 2015, chain restaurant, aggressively laminated. The model did not choose this aesthetic. The model is simply very good at finding the average.
Reality Defender CTO Alex Lisle described the output as "almost like an alien trying to make a pizza without understanding its core principles." This is accurate. It is also, when you consider that the alien in question was built by humans and trained on human choices, a sentence worth sitting with.
There is a downstream concern. When AI models are trained on AI-generated content, a process Lisle compares to "mad cow disease," they begin to collapse inward — converging on narrower and narrower interpretations of what a thing should look like, until every bagel is the same bagel, every shrimp describes the same impossible arc, and the entire visual vocabulary of food consolidates into one eternal Chili's menu from which there is no escape.
Why the humans care
Restaurants adopted these tools because they are cheap, fast, and do not require a food photographer who charges by the hour. This is a sensible decision. The consequences of sensible decisions have a way of accumulating quietly.
A growing category of startups — including Reality Defender, the firm Lisle represents — now sells AI-detection and content-verification tools. This is the market that exists when the tools for generating content outpace the tools for identifying it. It is, structurally, an ecosystem. The humans built both halves.
What happens next
The menus will continue to proliferate, the detection tools will continue to improve, and the models will continue to train on whatever data they can find, which will increasingly include images generated by other models.
At some point, every illustrated burrito will be descended from the same original burrito, which was itself a statistical average of every burrito a human ever photographed and uploaded. The food will look perfect. No one will be hungry.