AI-generated menu boards started appearing in delis, cafés and night markets this summer, and the internet reacted with something close to nausea. TechCrunch’s Amanda Silberling published the sharpest explanation of why on 3 September 2026, and her answer was not “the models are bad at food.” It was stranger and more useful than that: the images are unappetising precisely because they are too smooth, too symmetrical and too much like each other.
That last part is the interesting bit for anyone running a business rather than a restaurant. The revulsion people feel at an AI-generated menu is a consumer-facing symptom of a technical problem that also shows up in AI-generated copy, AI-generated logos and AI-generated code. Models trained to produce the average of everything eventually produce nothing but the average, and human beings are unnervingly good at spotting the average when it is pretending to be a specific plate of food.
This article walks through the reporting, the peer-reviewed psychology behind the disgust reaction, the feedback loop that makes each generation worse, the economics pushing small operators toward the cheap option, and the practical rules a marketing team should apply before letting a generated image anywhere near a customer. The same reasoning applies to any business using generative tools in customer-facing marketing services, not just to the ones selling lunch.
Table of contents
- What Actually Happened: AI-Generated Menu Boards Went Viral
- Why Every AI-Generated Menu Looks Like the Same Restaurant
- The Sameness Problem Has a Name: Convergence and Model Collapse
- The Science of Disgust: Why an AI-Generated Menu Turns Your Stomach
- The Feedback Loop That Makes Each AI-Generated Menu Worse
- Why Restaurants Reach for an AI-Generated Menu in the First Place
- What an AI-Generated Menu Costs You in Trust
- AI-Generated Menu Images and UK Advertising Rules
- The Sameness Problem Is Not Confined to Food
- How to Use AI in Food Marketing Without Shipping an AI-Generated Menu
- What to Do If You Have Already Published an AI-Generated Menu
- AI-Generated Menu FAQ
- References
What Actually Happened: AI-Generated Menu Boards Went Viral
The story broke the way most internet stories break now, from a photograph nobody planned to take.
A deli, a sign, and two million views
Business Insider’s Chris Murphy walked into his local deli in late August 2026 and found every menu item replaced by uncanny AI-generated menu images. He was not the only one. One X post about an AI-generated menu collected close to two million views, and the format spread fast enough that NBC News ran a piece on 1 September about the images jumping from timelines onto physical signage. The AI-generated menu pictures were memorable for the wrong reasons: shrimp that looked like life preservers, burrito fillings nobody could identify, protein with a bug-like texture over it.
The same reaction in four countries
What makes this more than a local oddity is how consistent the response was across markets. CNN documented AI-generated menu photography in Tokyo showing bagels with a cobweb texture, in Kuala Lumpur showing spaghetti marinara that looked like Play-Doh, and in Berlin showing quiche that read as plastic and pasta that read as rubber. Adaeze Onejeme, a 28-year-old ad sales worker who hit one at a Hollywood night market, put it plainly: “Those don’t look like chicken nuggets. Those don’t look like typical food.”
Nobody could quite say why
The striking thing in the reporting is how hard people found it to articulate the objection. Julia Damphouse, a 30-year-old tour guide in Berlin, said the images look “too geometric… like if you take acid and food stops looking like food.” Sam Biddle at the Intercept called an AI shrimp scampi “spiritually unsettling and cosmically unnerving.” Kat Kinsman, features editor at Food & Wine, described a “queasy upset.” These are professional writers reaching for metaphor because the literal description keeps escaping them.
The pushback arrived quickly
Rejection of the AI-generated menu trend came fast and physical. A San Francisco coffee shop spent hundreds of dollars removing graffiti that protested the AI-generated menu images on its facade. Jill Winger of the Chugwater Soda Fountain in Wyoming posted a handwritten sign promising the business would never use AI to promote itself, and it drew 119,000 likes. “I’m seeing which way the pendulum is swinging,” she said. Meanwhile a café customer discovered that an item advertised on the board did not exist at all, because ChatGPT had invented it.
Why Every AI-Generated Menu Looks Like the Same Restaurant
Alex Lisle, chief technology officer at the deepfake-detection firm Reality Defender, gave TechCrunch the line that explains the whole phenomenon: “A lot of this stuff looks like a Chili’s menu from 2015, and there’s a reason for that.”
The training data is a chain restaurant
The reason is that the corpus is full of chain-restaurant marketing photography. That material is abundant, well-lit, consistently framed and free of anything that would offend anyone, which makes it exactly the kind of data an image model learns most confidently. When a model is asked for a plate of food, it returns the centre of that distribution.
None of this is a defect in one particular tool. Decades of computer vision research would predict exactly this outcome from a corpus weighted so heavily toward a single style of commercial photography. The result is that an AI-generated menu in Wyoming and an AI-generated menu in Kuala Lumpur converge on the same visual dialect, because they are both drawing from the same statistical middle.
Optimisation for pleasingness produces homogenisation
Lee Rainie, who directs the Imagining the Digital Future Center at Elon University, described the mechanism: “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.” Safety tuning, aesthetic reward models and preference optimisation all push in the same direction. Each one shaves off outliers. The cumulative effect is that the model becomes very good at producing something inoffensive and very bad at producing something particular.
“Shaving off the edges”
Rainie’s phrasing for this is worth keeping: “What AI is known to do both in images and language is to shave off the edges.” Those edges are where food photography actually lives. A real croissant has an uneven bake. A real burger leaks. A real plate has a fingerprint on the rim and a shadow that falls wrong. An AI-generated menu image removes all of it, and removes with it the evidence a viewer uses to decide that a picture depicts one specific object rather than an idea of an object.
The alien making a pizza
Lisle’s other line captures the failure mode from the model’s side: “It’s almost like an alien trying to make a pizza without understanding its core principles.” The system has seen millions of pizzas and learned what a pizza looks like from the outside. It has never learned that cheese browns unevenly, that pepperoni cups when it cooks, or that the crust and the topping are made of different materials. So it produces a surface that satisfies every visual statistic about pizza while violating the physics a viewer knows without thinking.
| Detail | A real food photograph | An AI-generated menu image |
|---|---|---|
| Surface texture | Uneven browning, crumbs, moisture | Uniform sheen, plastic or rubber read |
| Portion edges | Irregular, sometimes spilling | Geometric, contained, symmetrical |
| Countable items | A fixed number you can count | Ambiguous count, merged shapes |
| Cutlery and props | Correct scale and reflection | Warped handles, impossible tines |
| Background | A real room with real clutter | Generic blur, no fixed location |
| Lighting | One light source, consistent shadows | Ambient glow with no source |
| Match to the kitchen | The dish the kitchen actually sends | Sometimes a dish that does not exist |
The Sameness Problem Has a Name: Convergence and Model Collapse
The homogenisation people are noticing on café walls is the visible end of a documented research problem.
What model collapse means
In 2024, Ilia Shumailov and colleagues published work in Nature showing that generative models trained on recursively generated data lose the tails of the original distribution and degrade irreversibly. The rare cases go first, then the merely uncommon ones, until the model can only produce the middle. Lisle described the same idea in blunter terms: “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.”
Convergence is the version we are actually living in
Full collapse is a laboratory endpoint. What the industry is experiencing now is the milder version, usually called convergence: output quality does not fail outright, it narrows. Every model trained on a web that increasingly contains its own previous output drifts toward the same aesthetic. That is why an AI-generated menu produced this month looks like one produced by a different tool in a different country, and why the visual signature is becoming recognisable to ordinary customers rather than only to specialists.
The web is now partly synthetic
The mechanism only works because generated images are being published at scale and then scraped back into the next training set. Menu boards are a small contributor, but they are a revealing one, because they are physical, photographed, uploaded and indexed. Each viral AI-generated menu photograph is a piece of synthetic imagery entering the corpus with a caption that describes it as food.
| Research | Sample or scope | What it found |
|---|---|---|
| Diel et al., Appetite, 2025 | 95 participants, 38 generated food images | Near-realistic images rated most uncanny and least pleasant |
| Diel follow-up study | Generated versus real food images | Lower desire to eat, identical nutrition judgements |
| Shumailov et al., Nature, 2024 | Recursively trained generative models | Distribution tails vanish; degradation is irreversible |
| Consumer survey, 2023 | General consumers | 60% could not separate generated from authentic food images |
The Science of Disgust: Why an AI-Generated Menu Turns Your Stomach
The strongest evidence that this is a real perceptual effect rather than internet snobbery comes from a peer-reviewed study published before the trend went viral.
The Duisburg-Essen experiment
Alexander Diel and colleagues at the University of Duisburg-Essen ran 95 participants through 38 AI-generated food images spanning a realism gradient, from obviously artificial through to near-photographic, including some depicting spoiled dishes. The paper, “Eerie edibles: Realism and food neophobia predict an uncanny valley in AI-generated food images,” appeared in the journal Appetite in 2025.
The uncanny valley applies to lunch
The finding is counter-intuitive and important. The most unsettling AI-generated menu images were not the obviously fake ones. They were the nearly-real ones. Participants rated slightly imperfect generated food as significantly more uncanny and less pleasant than either cartoonish images or highly realistic ones. That is the classic uncanny valley curve, previously documented for faces and robots, showing up for food. An AI-generated menu sits directly in the trough by design, because it is trying to look real and not quite managing it.
Food neophobia, not squeamishness
The study also identified who reacts hardest. People with higher food neophobia, the reluctance to try unfamiliar foods, experienced markedly more discomfort. General disgust sensitivity was not a significant predictor. In other words, the reaction is specifically about eating, not about images. Diel’s own summary: “Food is important to us… we are very sensitive to when something is even slightly off.”
It costs you appetite, not credibility
A follow-up study sharpened the commercial implication. Shown generated food images, people reported less desire to eat the dish while judging its nutritional qualities identically. The image did not make them think the food was worse. It made them not want it. For a business whose entire purpose is to make someone want the thing in the picture, that is the worst possible failure mode, and it is invisible in any metric short of sales.
Everyone feels it, few can name it
Rainie’s observation ties the research to the anecdotes: “People have an almost unexplainable sense about when they’re looking at something that’s AI-generated.” That inarticulacy is precisely why an AI-generated menu is dangerous for a small business. Customers rarely complain about the images. They just order less, or walk.
The Feedback Loop That Makes Each AI-Generated Menu Worse
There is a second, faster loop operating inside individual businesses, and it explains why these images get worse rather than better with effort.
Every revision smooths the image further
When an operator does not like an AI-generated menu image, the natural response is to feed it back and ask for a fix. Each pass applies the same averaging pressure to an image that has already been averaged. Detail that survived the first generation does not survive the fifth. The revision process, which in normal design work converges on something specific, here converges on something generic.
The hundred-iteration demonstration
An X user posting as Labtec ran the experiment to its conclusion, cycling an image through roughly a hundred generations. The verdict on the result was short: “The end result actually makes me uncomfortable.” That is the small-scale version of what the Nature paper describes at the level of a whole model, compressed into an afternoon and visible to anyone.
The chatbot will not tell you it is wrong
Daniel Silva, who runs the design firm Silva Heeren, identified the missing brake: “ChatGPT will tell the user that their design is genius.” A generative tool has no equivalent of a designer saying the layout does not work, the contrast fails, or the file is the wrong resolution for print. It confirms and produces. An operator with no design training gets encouragement at every step and no technical specification at any of them.
Print is where it falls apart
Silva’s other point is the practical one that catches most small businesses. Chat tools produce screen-resolution output with no bleed, no colour profile and no thought about the substrate. An AI-generated menu image that reads as merely odd on a phone reads as alarming when it is printed a metre wide, backlit, and hung over a counter where someone stands looking at it for two minutes while they queue.
Why Restaurants Reach for an AI-Generated Menu in the First Place
It is easy to mock the output and much harder to argue with the arithmetic that produces it.
The margins are genuinely brutal
Forty-two percent of restaurant operators reported an unprofitable 2025. Against that backdrop, a food photographer with a stylist, a half-day shoot and licensing is a discretionary line item that gets cut early. An AI-generated menu costs nothing and arrives in seconds. Paul Shovlin, an English professor at Ohio University who studies this, notes that the smallest and most budget-constrained establishments are exactly the ones adopting AI graphics.
It is usually an employee, not a strategy
The deli Chris Murphy visited had not commissioned an AI campaign. A member of staff made the AI-generated menu, and it went up as a stopgap while the business waited on a proper sign with real photographs. That is the typical shape of the decision: no procurement, no brief, no review, just someone solving a problem in ten minutes with a free tool. It is the same pattern that produces shadow IT everywhere else in a business.
Adoption is real but narrow
The National Restaurant Association’s 2026 State of the Industry report, released in February, found 26% of operators using AI-related tools, with marketing the single largest use case at 19% of full-service and 15% of limited-service operators. Only 10% used it for administrative work and 6% for customer orders. The technology entered restaurants through the marketing door, which is the one facing the customer.
Most operators think they are keeping pace
The same report found 60% of operators describing their technology use as roughly in line with competitors, 12% claiming a leading edge and 28% admitting they lag. That distribution matters, because it means the majority believe they are doing the normal thing. When the normal thing becomes an AI-generated menu, nobody involved experiences it as a risky decision.
What an AI-Generated Menu Costs You in Trust
The financial case for generated imagery is about the cost of production. The case against it is about the cost of everything downstream.
The picture is a promise
A menu photograph is not decoration. It is a representation of what arrives on the plate, and customers treat it as one. Once a business publishes an AI-generated menu, every image on it becomes a claim the kitchen cannot verify. The café whose ChatGPT-written board advertised an item that did not exist is the extreme version, but the ordinary version is subtler and more common: the dish arrives, it looks nothing like the AI-generated menu, and the customer quietly recalibrates how much of everything else to believe.
The credibility loss generalises
Lisle framed the wider stakes better than anyone: “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.” When a category of image stops being reliable, the doubt does not stay contained. This is the same erosion covered in our reporting on AI detection and whether it can be trusted, where the question of what counts as evidence of authorship has proved far harder than it looks.
Authenticity is the actual product
Nancy Hopkins, president of the International Association of Culinary Professionals, gave the advice that most food businesses already know and forget under pressure: “Tell your honest story. Real people, real stories will always be better.” A phone photograph of the real dish, taken badly in real light, outperforms a flawless AI-generated menu image on the only metric that matters, which is whether someone orders it.
Not everyone agrees it is a scandal
There is a reasonable counter-position and it deserves stating. Lyndsey Lozano, a coffee shop co-owner on the receiving end of the backlash, asked: “Is it really that horrific that you have to act on such a feeling?” Chris Murphy landed in a similar place: “I’m not saying it’s great that these places are using AI to create food monstrosities, but I’m saying I understand. If AI is going to erode society — and maybe it is — deli signage isn’t necessarily what I imagine will be the front lines.” Vandalising a small business over a sign is a disproportionate response to a bad sign.
| Force | Evidence from the reporting | Consequence |
|---|---|---|
| Margin pressure | 42% of operators unprofitable in 2025 | Photography is cut before rent or staff |
| Zero marginal cost | Image tools are free or near-free | No budget line means no approval step |
| No design review | “ChatGPT will tell the user that their design is genius” | Nothing catches resolution or scale errors |
| Informal ownership | Deli sign made by an employee as a stopgap | No brief, no sign-off, no removal date |
| Perceived normality | 60% believe their tech matches competitors | The decision never feels like a risk |
AI-Generated Menu Images and UK Advertising Rules
For a UK business the question is not only whether generated imagery is tasteful. It is whether it is compliant.
A menu photograph is an advertisement
The CAP Code applies to marketing communications, and a menu board outside a café is one. Its central requirement is that marketing must not materially mislead, including by exaggeration or by omission. A photograph that depicts a portion, a garnish or an ingredient the kitchen does not serve is a misleading representation regardless of whether a camera or a model produced it. The Advertising Standards Authority does not need a new AI rule to act on an AI-generated menu; the existing one already covers it.
The consumer protection regime sits behind it
Beyond the self-regulatory code, the consumer protection provisions now consolidated in the Digital Markets, Competition and Consumers Act 2024 give the Competition and Markets Authority direct enforcement powers over misleading commercial practices. A generated image that materially changes a customer’s decision to buy is squarely within that territory. The practical exposure for a small operator is small, but it is not zero, and it grows the moment the image implies something about ingredients.
Allergens and composition are the real hazard
The highest-risk case is not aesthetic at all. If a generated image shows sesame seeds on a bun that has none, or an ingredient a customer avoids, the picture is now interacting with food information rules rather than advertising taste. An AI-generated menu invents garnishes and toppings constantly, because inventing plausible detail is precisely what the model is optimised to do. That is a category of error a stock photograph does not make.
Provenance is becoming a labelling question
Disclosure regimes for synthetic media are tightening across jurisdictions, and the direction of travel is consistent: if an image is generated, the viewer is increasingly entitled to know. A business that adopts an AI-generated menu now should assume the labelling expectation is coming, and should be able to say which images are real.
The Sameness Problem Is Not Confined to Food
The AI-generated menu is simply the most legible example of something happening across every generated medium.
The same flattening hits written copy
The homogenisation Rainie describes applies identically to language. Generated marketing text converges on the same cadence, the same hedges and the same three-item lists, for the same reason: it is optimised toward an inoffensive centre. Readers detect it the way diners detect a generated photograph, without being able to name the tell. The parallel problem of separating machine-written from human-written work is covered in our piece on why distinguishing AI-assisted from AI-generated writing is much harder than a binary verdict suggests.
Brand differentiation gets harder, not easier
If every competitor in a sector uses the same tools with similar prompts, the outputs converge and the visual distance between brands shrinks. The businesses that gain from generative tools are the ones using them for volume work behind the scenes, not for the handful of assets that are supposed to say this business is not the others. Sameness is a competitive problem before it is an aesthetic one.
The tells will keep moving
It is worth being honest about the trajectory. Image models are improving quickly, and the specific defects that make today’s AI-generated menu obvious will not all survive the next two years. What is less likely to disappear is convergence, because it comes from the optimisation targets and the training corpus rather than from a fixable bug. A business planning around “the images will get better” is planning around the wrong variable. If you are tracking that trajectory, our AI models and tools hub follows each release and what it changes in practice.
How to Use AI in Food Marketing Without Shipping an AI-Generated Menu
None of this argues for banning generative tools. It argues for putting them where their weaknesses do not touch a customer’s decision to buy.
Never let a model depict the actual product
This is the one hard rule. Anything a customer will compare against a physical object should be photographed, not generated. A phone camera and a window is enough. The comparison is not between an AI-generated menu and a professional shoot; it is between an AI-generated menu image and the plate that arrives ten minutes later.
Use generation for the work nobody eats
There is a large amount of legitimate work available. Background textures, section dividers, social post backdrops, internal deck illustration, first-draft layout exploration, alt text, and translation of an existing menu into another language are all safe, because none of them makes a claim about the food. That is where the productivity actually is.
Put one human approval between the tool and the wall
The pattern behind almost every bad AI-generated menu is the absence of a review step, not the presence of a model. One named person who looks at the artwork at full size, in the light it will hang in, before it is printed, catches nearly all of this. It costs minutes.
Keep a record of what was generated
Note which assets were produced by a model and which were photographed. It takes a spreadsheet column. It answers a customer complaint, an ASA query and a future labelling requirement without an archaeology exercise, and it is the same discipline any business needs around AI use generally — the ground covered by a proper AI strategy rather than an ad-hoc one.
| Use | Main risk | Verdict |
|---|---|---|
| Photograph of a dish you sell | Misleading claim, appetite loss | Never generate |
| Ingredient or allergen close-up | Invented garnish, food information error | Never generate |
| Staff or premises imagery | Fabricated people and places | Never generate |
| Decorative pattern or divider | Low, aesthetic only | Safe with review |
| Layout and typography drafts | Print specification errors | Safe with a designer |
| Menu translation and alt text | Mistranslation of dish names | Safe with a native check |
| Internal decks and planning | Negligible | Safe |
What to Do If You Have Already Published an AI-Generated Menu
Plenty of businesses are now in this position, having made a reasonable decision that aged badly in about three weeks.
Take down anything that depicts a dish
Start with the AI-generated menu images that make a claim. A menu with no photographs is a perfectly normal menu and has been for most of the history of restaurants. Removing an AI-generated menu image costs nothing and immediately ends the mismatch between the picture and the plate.
Replace the top ten sellers first
You do not need a full shoot. Photograph your ten best-selling dishes on a bright day near a window, on the plates you actually use, and you have covered most of what customers look at. The imperfection is the feature: it is the evidence the dish exists.
Say what happened if anyone asks
Businesses that have handled this well were straightforward about it. The deli explained that a member of staff made the sign while a real one was on order. That is a completely acceptable answer and it defuses most of the criticism, because the objection is rarely to the business and usually to the sense of being shown something fake without being told.
Decide the rule before the next deadline
The reason an AI-generated menu appears is time pressure meeting a free tool. Write down the rule now, while nobody is under pressure: generated images never depict food we sell. A rule agreed in advance survives the next Friday afternoon; a judgement call made under deadline does not.
AI-Generated Menu FAQ
Why does every AI-generated menu look the same?
Because image models are trained on abundant, well-lit, inoffensive commercial food photography and optimised toward pleasingness, they return the centre of that distribution. Reality Defender’s Alex Lisle described the result as looking “like a Chili’s menu from 2015.”
Is the disgust reaction real or just internet snobbery?
It is real and measured. A 2025 study in Appetite with 95 participants found that near-realistic generated food images were rated significantly more uncanny and less pleasant than either obviously artificial or highly realistic ones.
Does an unappetising AI-generated menu actually cost sales?
A follow-up study found people reported less desire to eat generated food while judging its nutritional qualities identically. The image does not make customers think the food is bad; it makes them not want it, which is harder to detect and more expensive.
What is model collapse, and is it happening now?
Model collapse is the irreversible degradation described by Shumailov and colleagues in Nature in 2024, where models trained on their own output lose the rare cases first. What is happening now is the milder form, convergence: quality narrows rather than fails.
Will better models fix this?
Partly. The specific artefacts that make an AI-generated menu obvious today will improve. Convergence is less likely to improve, because it comes from the training corpus and the optimisation targets rather than from a bug.
Is an AI-generated menu legal in the UK?
An AI-generated menu is not itself prohibited, but a menu board is a marketing communication under the CAP Code and must not materially mislead. Images that misrepresent portions, garnishes or ingredients create exposure regardless of how they were produced.
What should a small food business do instead?
Photograph the dishes you sell, use generation only for decorative and internal work, and put one named person between the artwork and the wall. As Nancy Hopkins of the International Association of Culinary Professionals put it: “Real people, real stories will always be better.”
References
The sameness problem behind those unappetizing AI-generated menus
Appalling AI menu pictures keep going viral. But what makes them so revolting?
Side of slop? Bizarre AI food images jump from online to IRL menus
Delis are going viral for horrifying AI-slop menus
Eerie edibles: Realism and food neophobia predict an uncanny valley in AI-generated food images
AI models collapse when trained on recursively generated data
A note on Shumailov et al. (2024) on recursively generated data
National Restaurant Association State of the Industry
Over 25% of restaurant operators use AI
CAP Code: the UK Code of Non-broadcast Advertising and Direct & Promotional Marketing
Digital Markets, Competition and Consumers Act 2024
Food labelling: giving food information to consumers