AI-generated restaurant menus have converged on a single visual dialect — glossy, symmetrical, faintly plastic — and diners are starting to recoil from them before they can articulate why. Reality Defender CTO Alex Lisle traces the effect to the training corpus itself: chain-restaurant menus from the mid-2010s, reproduced and reinforced by every generation of image model since. An X user named Labtec ran the pattern to its extreme, editing a single ChatGPT-generated menu 100 times and watching the food warp into something viscerally wrong.
The finished sequence, Labtec wrote, "actually makes me uncomfortable." Each pass sanded another edge off the bagels, rounded another scoop of ice cream, smoothed another burger bun into a cartoon dome. TechCrunch replicated the experiment and reported similar decay.
The mechanic is not mysterious. Large language models and diffusion models like ChatGPT and Midjourney predict what a user wants by pattern-matching against training data, and the training data for "fast food menu" is dominated by Wendy's, Burger King, and McDonald's. Those three chains already share a visual language. The models absorb that language, output more of it, and the outputs feed back into the next scrape.
Key facts
- 01A ChatGPT experiment by X user Labtec edited the same AI-generated restaurant menu 100 times, producing progressively smoother, more uncanny food images.
- 02Reality Defender CTO Alex Lisle attributes the effect to convergence — a milder cousin of model collapse — as models train on their own outputs.
- 03Researchers at the University of Duisburg-Essen in Germany found AI-generated food images trigger a stronger 'uncanny valley' disgust response than obviously fake ones.
- 04The aesthetic default traces to fast-food chain menus from Wendy's, Burger King, and McDonald's — the corpus most heavily represented in training data.
Lisle called the base aesthetic a Chili's menu from 2015 and pointed out that this was the corpus from which the models drew their function. That is the entire explanation for why an independent bagel cafe's AI menu looks like it was generated in a suburban strip mall.
The word Lisle uses for the degradation is convergence, not model collapse. Collapse is the terminal case — a model fed too much of its own output eventually breaks. Convergence is milder: the model still works, but its range narrows, its outputs homogenize, and the median image drifts toward whatever the training set overrepresented.
Amazon has been sourcing rare books to scan into training data and destroying the physical copies afterward, a sign of how valuable fresh, non-synthetic material has become. As the open web fills with AI-generated content, the marginal cost of clean data keeps rising, and the marginal risk of convergence keeps rising with it.
Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, frames the same problem from the demand side. Models are tuned to produce outputs that are pleasing and inoffensive, which means the tails of the distribution — the weird, the specific, the regionally distinct — get cut off. Every menu ends up optimized toward the same safe middle.
There is also a physiological reason diners flinch. Researchers at the University of Duisburg-Essen in Germany found that AI-generated food images produced a stronger disgust response when they looked almost real than when they looked obviously fake. The uncanny valley applies to hamburgers.
Rainie argues that people have an almost unexplainable sense for when they are looking at something AI-generated, and that this instinct is why the backlash against AI restaurant menus has been so pronounced. Restaurants adopting the tools to cut design costs are discovering that customers read the aesthetic as a signal about the food itself.
The wider implication Lisle flags goes beyond lunch. Court systems, he notes, are built on the premise that taped confessions and videotaped evidence are the gold standard — that seeing and hearing is believing. Generative models have made that premise negotiable, and the shift, in his phrasing, is fundamental, for good or for ill.
The menu problem is a canary. Convergence is easy to spot when the subject is a burrito because everyone knows what a burrito looks like, and everyone can tell when a burrito looks wrong. In domains where the reader has no ground truth — legal briefs, medical summaries, financial explainers — the same smoothing is happening without the visceral tell. The next round of frontier models will either find ways to price fresh human data properly and preserve range, or the entire generative economy will keep drifting toward a single glossy mean that nobody actually ordered.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
