Bacon and eggs in a glass of cola: why AI menu photos need the restaurant's sign-off
Just Eat paused its AI menu images after soda bread turned into soda and a cheese roll became sushi. The model wasn't the only problem. If a platform makes pictures of a seller's product, the seller should see them before customers do.
In October 2026 Just Eat apologized and paused a trial of AI-generated images for menu items that had no photos, after the tool misread dish names, turning bacon and egg soda bread into food floating in cola and a cheese roll into sushi. The lesson for any marketplace generating images for sellers: local and ambiguous names will trip the model, so route every generated image to the seller for approval, let them reject it easily, and fall back to no image rather than a wrong one.
What happened
Just Eat, the UK food delivery platform, has been trialling AI-generated pictures for menu items that didn't have a photo. The idea is sensible on paper. A dish with a picture usually does better than a line of text, and plenty of small takeaways never got round to photographing their whole menu.
Then the pictures showed up on the app. According to the BBC and The Independent, a Belfast cafe's bacon and egg soda bread appeared as a glass of fizzy drink with a fried egg and bacon floating in it. A Manchester "chip muffin", which is a chip butty to most people, became chips poking out of a baked muffin. A misspelling of "plain" produced garlic bread shaped like an aeroplane. A cheese roll turned into a sushi roll. Salt and pepper chips, normally fried with onions, garlic and chillies, came out as plain-looking potatoes.
People shared the images widely. Just Eat apologized, said it was removing the inaccurate images, and paused the rollout while it develops the tool further. Its spokesperson also pointed out that the company works with tens of thousands of small restaurants on their food imagery, including funding local photographers.
The images were labelled as AI-generated. That didn't help much. One customer quoted by the BBC put the real issue simply: "I want to see what the dish actually looks like as it helps me decide who to order from or what to order."
What went wrong, as far as we can tell
We don't know how Just Eat's pipeline worked internally, so this is a reading of the public results, not an inside account. Still, the errors fall into a clear pattern.
Local names. Soda bread, chip muffin and salt and pepper chips are perfectly clear to people in Belfast, Manchester or anyone who orders from a Chinese takeaway in the UK. To a general-purpose model they're ambiguous, and it picked the wrong meaning with complete confidence.
Spelling as written. Menus on delivery apps often come from a restaurant typing quickly into a form. "Plane garlic bread" is a typo any human reader corrects without noticing. The model drew it literally.
No one between the model and the customer. This is the part that matters most. Each wrong image was the kind a restaurant owner would have rejected in a second. The BBC reported that many of the odd images were still live on menus when it looked. Whatever review existed, it didn't include the one person guaranteed to know what the dish looks like.
That last point applies to every marketplace, not just food delivery. If you generate a picture of someone else's product, you're making a claim about their product, on their listing, under their name.
Decision guide
Where each error came from
| Consideration | What the menu said | What the AI showed |
|---|---|---|
| Local name | Bacon and egg soda bread. | Bacon and egg floating in a glass of cola. |
| Regional dish | Chip muffin (a chip butty). | Chips baked into a muffin. |
| Typo | Plane (plain) garlic bread. | Garlic bread shaped like a plane. |
| Ambiguous word | Cheese roll. | Cheese rolled into a sushi roll. |
A missing photo costs you a little conversion. A wrong photo costs you the customer's trust in every other photo on the page.
What a better flow looks like
Here's the approach we'd suggest to any marketplace considering generated images for seller listings, whether that's meals, handmade goods, rentals or secondhand items.
Generate as a draft, not a listing. The image goes into a queue for the seller, not onto the page. The listing keeps showing what it showed before, which is usually no photo, until the seller says yes.
Make yes and no equally easy. One tap to approve, one tap to reject, and a short field for "what's wrong". If rejecting takes more effort than approving, busy sellers will approve whatever is there.
Show the seller what the model was told. Display the item name and description the image was generated from. When the seller sees "plane garlic bread" next to a picture of a plane, they fix the typo, and the next image is better too.
Default to nothing. If the seller doesn't respond within some period, the image stays unpublished. A missing photo costs you a little conversion. A wrong photo costs you the customer's trust in every other photo on the page.
Give customers a report button. Even approved images will sometimes be wrong. A simple "this doesn't look like what I got" link on the item lets the problem reach someone quickly.
How the work moves
A generated image, seller first
- Generate draft
Image goes to a seller queue, not the listing.
- Show the input
Seller sees the item name the image came from.
- Approve or reject
One tap each, with an optional reason.
- Default to none
No answer means nothing is published.
- Customer report
A link on the item if it still looks wrong.
The business case for slowing down
A seller approval step looks like friction, and it will reduce how many images go live in the first month. That's the point. The trial at Just Eat apparently aimed to fill gaps on menus at scale. When some of the results were wrong in obvious ways, the whole program had to stop, including the images that were fine.
An approval step also protects the marketplace's relationship with sellers. A restaurant that finds a picture of floating bacon on its own listing from a customer screenshot has every reason to be angry. A restaurant that gets a draft and says "not quite, the soda bread is a flat round loaf" is helping you build a better catalog.
There's a data benefit too. Every rejection with a reason is a labelled example of where the model gets local language wrong. Over time that becomes a list of terms to handle carefully: regional dishes, brand names, slang, words with two meanings. You can't get that list from a model provider. Only your sellers have it.
Finally, think about who pays for photography when the AI route is paused. Just Eat already funds local photographers for some restaurants. A marketplace that offers sellers a choice between a quick generated draft and a booked photo session gives them a real decision, rather than a picture that simply appears. Sellers who care about how their food or products look will pick the photo. The rest can approve the draft, knowing what they approved.
Where AI images do fit
None of this means generated images have no place on a marketplace. They can be useful for category headers, empty states, or generic items where nobody expects an exact match. They can also help a seller who wants a cleaner background on their own photo. In each case the seller starts from something they already know is accurate.
The line is easy to state. If the customer will reasonably assume the picture shows the actual thing they're buying, it should be the actual thing, or something the seller has looked at and approved. A label saying "AI-generated" tells customers the picture is not a photo. It doesn't tell them whether the picture is accurate.
If you run a marketplace and want generated content, seller approvals and listing quality designed together, that's what our marketplaces team does. Talk to us.
Related work: Nine's Style.
We build alongside the people doing the work.
Bring us the problem you’re thinking about.