Just Eat has paused the rollout of an AI image generator that was filling in missing dish photography across its marketplace, after customers in the United Kingdom began flagging results that ranged from the baffling to the biologically alarming. The clearest example came from a café in Belfast: its bacon-and-egg soda bread was rendered as a fried egg and rashers floating inside a glass of fizzy cola. The system read “soda” and committed entirely to the beverage.
- Not a rendering failure: Nearly every documented error stems from misread regional vocabulary, typos or homonyms rather than poor image quality.
- Paused, not cancelled: The company says it has suspended the rollout while it keeps developing the tool, and is pulling only the incorrect images.
- Little transparency: No figures on affected images, restaurants or countries have been disclosed, and the underlying model has not been named.
- Restaurants absorb the damage: The listings publicly misrepresented were ones owners had not commissioned images for.
A vocabulary problem, not a rendering problem
The pattern holds wherever a dish has a local name. A chip muffin — in parts of Manchester, a sandwich filled with fries — came back as a baked potato stuffed inside a muffin. A cheese roll, a bread bun with cheese, was depicted as sushi-style rolls of sliced cheese. Salt and pepper chips, a Chinese takeaway staple with onion, garlic and fried chilli, arrived as plain fries. Cheddar listed in a café’s vegetarian section turned into an orange fondue-style sauce, and burger sauce became a miniature hamburger sitting inside a small pot. Elsewhere, a jacket potato reportedly became a winter jacket covered in potatoes, and tiger prawns became a bowl of tiger cubs in sauce.
The mechanics explain the mess. An image generator has no knowledge of any specific dish; it draws the statistical average of everything labelled with those words. When a name is generic, the average lands close to reality. When it is a nickname, a homonym or a typo, the model nails the form and misses the referent. The neatest illustration may be a garlic-bread loaf shaped like an aircraft — the result, according to reporting, of “plain” being read as “plane”.
The more consequential risk is duller than floating bacon. Absurd images denounce themselves; nobody orders a cola with breakfast. Plausible images do not. A dish that appears correct but with two pieces instead of four, or with a side that is never delivered, ships quietly, disappoints on arrival and leaves a review for the kitchen rather than the algorithm.
“I want to see what the dish actually looks like as it helps me decide who to order from or what to order.” These “perfect” AI versions “only lead to disappointment if that’s what you think your food is going to be.”
What the company has actually done
Just Eat’s response is narrower than some coverage implied. An unnamed spokesperson said the company is “in the process of removing” the images and has “paused the rollout while we continue to develop the tool”. It is a suspension, not a cancellation, and only the incorrect renders are being taken down. The full statement apologises “for any inaccurate image that may have been published” and frames the episode as normal iteration: “as with any new technology, it doesn’t always go entirely right the first time.”
The company says it works with tens of thousands of small restaurants on product imagery, including funding local professional photographers — the only figure it has offered, and one covering the wider programme. Each generated photo does carry an on-app label marking it as AI-created, a genuine point in the system’s favour. Against that: reporters checking on 21 October found many of the most outlandish images still live, and Just Eat has not identified which tool or model it used, nor when the pilot began. Just Eat has also said that some examples circulating online were not produced by its tool, leaving the true error count unclear.
The economics behind the shortcut
The motivation behind the experiment was scale, not laziness. Professional food photography is expensive; generating an image is close to free. With tens of thousands of small partners, the arithmetic is obvious — and the experiment shows precisely where the saving breaks down, which is language rather than pixels.
Every source places the trial in the United Kingdom, and one Dutch outlet specifies that it ran only there. The brand has operated in Spain under the same name for more than a decade and strengthened its position by acquiring rival La Nevera Roja, but the group has not said whether an international rollout was planned. Detection, notably, came from customers — a UK subreddit and an Instagram account — rather than internal quality control. No restaurant association, consumer group or regulator has publicly reacted; the UK advertising authority warned in February 2025 that misleading imagery falls under its codes however it is produced, but that guidance names neither delivery apps nor this company.
The labelling is the part that worked. Without it, the reputational cost of a bacon-and-egg cola would have settled on a Belfast café rather than the platform that conjured it. Any next iteration will need two things the pilot lacked: sign-off from the restaurants whose menus are being illustrated, and enough grounding in regional food culture to know that soda bread is bread. The bigger lesson is that synthetic content scales cheaply right up to the point where it meets a community’s own words for its own food.
Sources & References