Just Eat pauses AI menu-image rollout after breakfast food appears in cola

Breakfast ingredients floating in cola were supposed to illustrate sandwiches. Just Eat has apologized for inaccurate AI-generated menu pictures and says it is removing them while pausing the tool’s rollout. The reversal puts a practical limit on its experiment: filling an empty photo slot is not enough if the picture shows the wrong meal.

Soda bread becomes a soft drink

The trial used AI to generate pictures for restaurant menu items that lacked photographs. At a Northern Ireland restaurant selling soda breads, the results included sausages, bacon and fried eggs floating in fizzy drinks, rather than pictures of the breakfast sandwiches on offer. TechRadar described the example in its October 20 coverage.

Shattered’s October 24 account, citing The Independent and other coverage, suggested the tool had interpreted “soda” as a soft drink instead of part of the bread’s name. That is a reported explanation, not a confirmed technical diagnosis. It points to a specific mismatch between the menu’s wording and the resulting illustration.

The same account listed other inaccurate pictures, extending the problem beyond the breakfast example:

• A miniature burger pictured inside a sauce pot.
• Sushi rolls shown with cheese inside.
• Wine glasses containing chunks of watermelon.

An apology and a halt to the rollout

Just Eat acknowledged the errors in a statement to The Independent, reproduced by TechRadar. Its spokesperson described the experiment as a trial of “AI-enhanced images” and apologized for inaccurate pictures that had been posted. The company presented the pause as time to develop the tool further, rather than an announcement that it was abandoning the approach.

We're in the process of removing them and have paused rolling this out while we further develop the tool to ensure it best serves our partners and customers.

Just Eat spokesperson, speaking to The Independent, as quoted by TechRadar

The wording leaves an important distinction. Just Eat called the pictures enhanced, while the trial was described as creating images where no photograph existed. This was not simply a dispute over how much an existing food photo had been polished: the reported purpose was to supply a visual for an otherwise unpictured item.

A label does not settle accuracy

Metro reported that the app identified the images as AI-generated, according to Shattered’s account. The reported label disclosed how the pictures were made; it did not prevent the inaccurate depictions that prompted the apology. Disclosure and a faithful representation of the dish are separate issues in this trial.

Michael Bortinger, Tastewise’s director of marketing and partnerships, offered a related warning in TechRadar’s coverage. He argued that an image model produces an average of images associated with a dish’s name, rather than a representation of one restaurant’s specific dish. Better-looking output, in his view, can make that average more convincing without resolving the underlying mismatch.

Just Eat described the number of inaccurate images as “incredibly small,” Shattered reported, but supplied no exact figure or count of affected restaurants. That leaves the scale unresolved. The striking examples show what went wrong, but do not establish how often the tool failed across the trial.

For any further rollout, the unresolved issue is how pictures will be checked against the food actually sold. Just Eat’s statement did not explain whether restaurants had approved the images before publication. Its commitment is to further development; the statement does not describe a revised approval process for getting those pictures back onto menus.