AI-generated restaurant menus are triggering widespread disgust because image generators like Midjourney and ChatGPT are trained on datasets that reinforce an eerily perfect, homogenized aesthetic. When customers see these menus, they experience an "uncanny valley" effect, where food that looks almost real but subtly wrong elicits more unease than obviously fake images. The problem reveals a fundamental limitation in how large language models (LLMs) and diffusion models learn from data.

Why Do AI Food Images Look So Unnaturally Perfect?

Walk into a cafe and you might notice something off about the bagel sandwich illustrations on the menu. Each one looks flawlessly symmetrical, impossibly smooth, and oddly appetizing in a way that feels wrong. Sometimes the results are egregiously fake, like a burrito with cheese so bubbly and melty it resembles avant-garde art rather than lunch. Other times, the images are so ordinary that you only notice something is amiss when you look closely.

The root cause lies in how these AI models are built. LLMs and diffusion models are trained on vast quantities of data, then identify patterns to predict what users are asking for. When someone requests "Make me a menu for a burger restaurant," the model references the most common burger restaurant menus in its training data. "A lot of this stuff looks like a Chili's menu from 2015, and there's a reason for that," explained Alex Lisle, Chief Technology Officer at Reality Defender, a company that develops AI-detection tools. "That was the corpus of work from which [the models] drew their function".

"It's almost like an alien trying to make a pizza without understanding its core principles," said Alex Lisle, Chief Technology Officer at Reality Defender.

Alex Lisle, Chief Technology Officer at Reality Defender

What Is Convergence, and How Does It Degrade AI Outputs?

The phenomenon at work here is called "convergence," which is distinct from but related to "model collapse." Model collapse occurs when AI models train on too much of their own AI-generated content, causing the system to degrade severely, like a form of digital inbreeding. Convergence is less extreme but still problematic, degrading output quality without making the model entirely useless.

Here's how the cycle works: popular restaurant menus already share similar visual styles. When an AI model generates a menu based on these sources, it mimics that same style. If that AI-generated menu ends up back in the training data for future models, it reinforces the homogenized aesthetic even further. Each iteration smooths out the edges and pushes toward a narrow band of "pleasingness".

Lee Rainie, director of the Imagining the Digital Future Center at Elon University, noted that this effect extends beyond food imagery. "The optimization of the datasets is for pleasingness, or you know, not being offensive, and so there's a way that turns into homogenization," Rainie explained. "What AI is known to do both in images and language is to shave off the edges".

"People have an almost unexplainable sense about when they're looking at something that's AI-generated, compared with something that was real in the first place," said Lee Rainie, director of the Imagining the Digital Future Center at Elon University.

Lee Rainie, Director of the Imagining the Digital Future Center at Elon University

How to Spot the Telltale Signs of AI-Generated Food Images

• Excessive Symmetry: Real food is messy and asymmetrical. AI-generated food often features perfectly round ice cream scoops, evenly distributed toppings, and mirror-image halves that don't occur in nature.
• Unnatural Smoothness: Textures in AI images tend to be overly polished and glossy. Cheese appears unnaturally melted, bread lacks realistic crust variation, and surfaces lack the imperfections of actual food photography.
• Anatomical Impossibilities: Some AI-generated seafood exhibits bizarre features, like shrimp that appear to eat their own tails, creating what researchers call "Lovecraftian food horrors."
• Homogenized Aesthetics: Multiple items on the same menu look like they were designed by the same hand, lacking the visual diversity of real restaurant photography or hand-drawn illustrations.

A user on X named Labtec conducted an experiment that illustrates this degradation. They created a restaurant menu in ChatGPT, then edited it 100 times to see how the food images changed with each revision. With each small edit to prices or item names, the food became progressively rounder and smoother. "The end result actually makes me uncomfortable," Labtec wrote.

What Does Science Say About Our Disgust Response?

Our aversion to these images isn't purely psychological. Researchers at the University of Duisburg-Essen in Germany found that AI-generated food images exhibited an "uncanny valley" effect, where images of food that looked almost real elicited more disgust and unease than images that were obviously fake. This is the same phenomenon that makes hyper-realistic humanoid robots feel creepy rather than comforting.

The cultural context amplifies this reaction. As AI-generated content becomes more visible in everyday life, people are developing an intuitive sense for detecting it, even when they can't articulate exactly why something feels off. Rainie emphasized that this sensibility is real and measurable: "There's just a sensibility that people sometimes find hard to articulate, 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] is so pronounced".

Rainie

The implications extend far beyond restaurant menus. Lisle warned that the shift toward AI-generated content has broader consequences for how we evaluate evidence and truth. "Seeing and hearing has always been believing, to the point where even our court systems are entirely tuned to the idea that the gold standard in evidence is taped confessions and videotaped evidence," Lisle stated. "That's no longer the case. The world has fundamentally shifted, for good or for ill".

Lisle

For restaurants currently using AI-generated menus, the message is clear: if customers react negatively to these images, that's probably reason enough to stop. The convergence problem reveals a fundamental challenge in AI development: models trained on curated, optimized datasets will inevitably produce outputs that feel sterile and homogenized, no matter how sophisticated the underlying technology becomes.