Just Eat Pauses AI Menu Image Trial After Food Misfires

When a menu image gets the food wrong

Just Eat has put its trial of AI-generated menu images on hold after the system produced pictures that misrepresented familiar British dishes. The company said it was removing the images while it worked to improve the technology. The pause brings a basic problem into focus: a picture on a food ordering service isn't decoration alone. Customers may use it to judge what a dish contains and what portion or style to expect, so an image that changes the food can make a practical tool misleading.

Among the results, a jacket potato appeared as a puffy coat decorated with potatoes. Soda-bread sandwiches became glasses of cola filled with bacon, sausages and fried eggs, according to Dangerous Minds. The examples are comic, but the underlying failure is plain. A system can produce a recognizable collection of food-related shapes while missing the meaning of the dish it was asked to depict.

Why images matter to ordering

Restaurants without menu photos may see generated imagery as a way to make listings more visually complete. A photo can help customers distinguish dishes, notice ingredients, and form an expectation before placing an order. Yet an image only serves that purpose if it corresponds to the item described. If it invents a different presentation or combines ingredients in an implausible way, it may leave customers less informed than a text description alone.

That gap matters because customers generally can't inspect a delivered meal before deciding to buy it. They rely on the menu's words and pictures, alongside whatever other information the service provides. A generated image that misstates the dish can therefore shape a choice on a false premise, even when no one intended to mislead. This is more significant than it looks: the image is part of the product description.

There are limits to what a menu image can promise. A restaurant's actual serving may vary in appearance, and a photo can't communicate every detail of preparation or portion size. But variation is different from depicting a jacket potato as clothing or turning sandwiches into drinks. The first is an ordinary gap between a representation and a meal; the second suggests the system hasn't reliably understood the item.

The challenge for generated imagery

Image-generation systems can create convincing-looking pictures without reliably preserving the specific meaning of a prompt. Food names can be local, compound, or culturally familiar; the software may associate a word with its separate components rather than the dish those words describe together. In this case, the reported results suggest that the system mishandled British food references. They don't, by themselves, establish how often the problem occurred or whether other dishes were affected.

That distinction matters. A striking error can reveal a weakness, but it can't tell readers how common the weakness is across a full menu or service. Nor does a plausible-looking image prove that the pictured dish reflects a restaurant's recipe. Similar systems typically depend on review and correction when the output is meant to represent a specific real-world item. The closer the image comes to a purchasing decision, the less safe it is to treat plausible appearance as evidence of accuracy.

Here's the thing: a tool intended to fill a missing-photo gap can create extra work if each generated image needs careful checking. Someone has to compare it with the menu item and decide whether it represents the dish faithfully. If the system's output is inconsistent, the restaurant or platform may face a choice between leaving a listing without an image and using one that risks confusing customers. Neither outcome delivers the simple fix the technology is meant to offer.

Review is part of the product

The company says it's removing the images while it works to improve the technology. That response acknowledges that the current results weren't suitable for display, but the statement doesn't specify what changes are planned or how the company will assess whether the images are accurate. Those details would help explain how the trial might resume. Without them, the central question remains operational: what level of human review will be required before a generated picture appears beside a dish?

Review needn't mean that every image is assessed by a specialist in food or technology. It does mean someone should be able to check the image against the actual menu description and flag obvious mismatches before customers see it. Restaurants may be best placed to confirm whether an image resembles what they serve, while the platform controls how listings are presented. That division of responsibility can be awkward, particularly if businesses are asked to review a feature intended to save them time.

(And this is the part most people overlook.) Accuracy isn't just a question of whether the image looks appetizing. It has to communicate the right dish. A polished but incorrect picture can be more persuasive than a visibly rough one, which makes review important even when the output looks professional at first glance.

What restaurants and customers should take from the pause

For restaurants, the takeaway is practical: AI-generated visuals need close review before they stand in for the food customers are ordering. If you're working in this space, the relevant test isn't whether a picture looks good in isolation; it's whether a customer could match it to the listed dish without being misled. Text descriptions and a missing image may be preferable while a generated alternative remains uncertain. A picture that gets the food wrong.

For customers, the episode is a reminder that a menu image can be a representation rather than a direct photograph of the meal. That doesn't make every generated image unreliable, and the source doesn't establish how broadly this trial's errors appeared. It does show why visual presentation needs to be judged alongside the written menu. When the two don't agree, the image may be offering confidence the listing hasn't earned.

Implications for the trial

Putting the trial on hold gives Just Eat an opportunity to address the mismatch before the images continue appearing on menus. The episode also sets a clear standard for any return: generated pictures should help customers recognize what they're ordering, and restaurants should be able to identify when they don't. Better image generation may reduce some errors, but the source provides no evidence yet about what improvements are underway or whether they will solve this category of mistake.

The broader lesson is modest but useful. AI can produce an image quickly; that doesn't make the image dependable as a description of a particular dish. For a food-ordering service, the value of the feature will depend less on how many listings gain pictures than on whether those pictures match the food. Until that match is dependable, the technology's convenience comes with a review burden the trial was meant to ease.