By asahifile 
Just Eat has apologised after its AI-generated food photography became a running joke. The delivery platform was testing the technology to illustrate dishes that had no pictures yet, and the model produced bacon and eggs floating in a glass of cola, chips poking out of muffins, a cheese roll rendered as sushi, and garlic bread shaped like an aeroplane.
The company said it is removing the incorrect images and will stop generating more “until the AI was further developed”. It is a small, funny story. It is also a tidy demonstration of the failure mode that keeps catching businesses out.
The bug is semantic, not visual
Every one of these errors comes from the same place: the model resolved ordinary words by statistical habit instead of local meaning. “Soda” in Belfast means soda bread, so a bacon and egg soda became a fizzy drink with a fried egg in it. A “chip muffin” in Manchester is a chip butty, so the model baked potato into bread. Garlic bread went airborne because a menu typo read “plain” as “plane”.
None of that is a rendering defect. The pictures are sharp, appetising and plausible; they are simply about the wrong food. When a system is asked to turn a short noun phrase into an image, it has no way to know which of the phrase’s meanings applies to a specific takeaway on a specific street, and it will not ask. It will answer confidently instead.
Why companies reach for this anyway
The temptation is real and rational. Food photography is expensive, and most small restaurants have no usable images of their own. Just Eat funds professional local photographers and says it works with tens of thousands of small restaurants on their imagery — a commitment that cannot cover every menu item on every menu.
So there is a genuine gap, and generated images look like free coverage of it. But a product photo is not decoration. It is a claim about what will arrive in the bag. A wrong picture is therefore not an aesthetic error; it is a false statement about a physical good, made at scale, on the page where a customer decides whether to trust the seller.
The label does not fix it
To its credit, Just Eat labels AI-generated images. Transparency is the right instinct. But a disclosure does not make an inaccurate picture accurate; it only moves the harm from deception to disappointment. And the commercial incentive runs the other way: the whole point of improving these models is to reach the point where nobody needs to check whether an image was generated. That is precisely the moment consumers stop being able to spot the wrong concoction at all.
There is a broader pattern here. The same class of error shows up whenever automation is applied to the last mile of context — a summary that gets a number right but the subject wrong, a translation that keeps the words and loses the register, a support bot that answers the question it recognised rather than the one that was asked. Fluent output makes these mistakes harder to notice, not easier.
There is also a mundane explanation for why this keeps happening, and it has nothing to do with the technology being new. Generation is cheap; review is not. The cost that disappears when a model produces a picture is the cost of the person who would have looked at it and said that is not a chip muffin. Putting that person back is the entire fix — and it is usually the first thing an efficiency programme removes.
What to watch
- Whether Just Eat restores human review before images go live, rather than reviewing after a subreddit notices.
- Whether advertising regulators in the UK or EU treat AI-generated product imagery as a misleading commercial practice when it misrepresents the goods.
- Whether “AI-generated” labels become a standard requirement on product pages, and whether platforms keep them once models improve.
- How quickly the same test lands on other categories — groceries, marketplaces, travel and property listings — where a plausible image can misrepresent something a buyer cannot inspect first.
Sources: BBC News.