Just Eat has paused the rollout of an AI tool that generates menu photos, after the trial put bacon and fried eggs in a glass of cola.
The images appeared on the listing for a Northern Ireland restaurant that sells soda breads. Instead of the Irish breakfast sandwiches on the menu, customers saw sausages, bacon and fried eggs floating in pints of fizzy drinks. Other botched images circulated on social media, and one Just Eat user wrote on X on October 9 that if their salt and pepper chips arrived looking like that, they would "cause murder".
A Just Eat spokesperson said the company had been trialling "AI-enhanced images", was removing the inaccurate ones and had paused the rollout while it develops the tool further.
The company has company. DoorDash has opened a text-to-order AI agent for testing: you text what you want, it picks a dish and texts back a photo. Journalists testing it found several errors, including photos that did not match the recommendation.
The cola sandwich is the cheap kind of failure. It is absurd, so it gets screenshotted, shared and pulled within days. The harder case is the image that looks plausible and is wrong in a small way, such as a wrong-sized dish or a missing topping. Nobody posts those, so nobody fixes them.
Michael Bortinger, director of marketing and partnerships at the food-data firm Tastewise, gave the technical reason: "An image model has no concept of a specific dish. It renders the average of everything labeled 'burger', and better tools just make that average more convincing." In other words, improving the model makes the errors harder to spot, not rarer.
The appeal is easy to see. Photographing every item is slow, and a generated image costs a restaurant no time. But a menu photo is a promise about what arrives at the door, and a picture of the average sandwich does not keep it. The savings go to platforms and restaurants, while customers carry the risk of a mismatch. If a generated photo overpromises and the error is subtle, who is supposed to notice?