A breakfast that became a drink
Just Eat spent months trying out an AI tool that makes menu images for dishes with no photograph. When the results reached customers, a Northern Ireland cafe's soda bread breakfast was illustrated with a pint of cola with a meal floating in it. A scampi listing showed sushi. Salt and pepper chips, which usually come with fried onions, chili and peppers, appeared without them.
The company apologized, said new technology "isn't always quite right the first time," removed the images and paused the rollout. It had also told The Journal that every AI image is manually checked by a person. That statement was about the trial in general. Nothing published says how the cola image got through, or whether anyone looked at it.
A review that approves every error
Look at the pictures as a reviewer would. Each one is well made. A reviewer asking whether it is a good photo of food has no reason to object, and a rollout built on that question would pass all of them. The mistake is only visible against a different question: is this the dish we actually serve? Only the cafe can answer that.
Just Eat has not said what went wrong inside the tool. The pattern fits a model that read a word by its everyday meaning, but that is an inference from the output, not a finding.
Every team has a menu
Inside an organization, ordinary words pick up local meanings. A ticket is "resolved" when the customer confirms in one department and when it is closed in another. A "priority" is an order, a label or a deadline depending on who is speaking. A person new to the team learns these in weeks. A model has no weeks, and without help it falls back on the common meaning and produces something that reads as competent.
A house-terms sheet
The fix is dull and it works. Three steps:
Write the sheet. List the dozen or so words your team uses differently from outsiders. For each, give the everyday meaning, the meaning here and one example.
Attach it to the brief. Put the sheet in the same request as the task, every time, so the tool never has to guess. Then ask it to list any term it still had to assume.
Give it one owner. Someone who knows what the words mean keeps the sheet current and reviews anything customer-facing. Review by whoever happens to be free tests whether the output looks right, and that is the question that passed the cola.
A sheet works best when the people using it have practiced briefing with it: Learn to brief AI with your team's own terms →
Keep reading: In AI, Context is King · Ask AI what you left ambiguous
This is the thinking behind a Trusted AI Culture.
Sources: The Independent, "JustEat apologises over AI menu pictures including bacon and eggs floating in cola," October 19, 2026. The Journal, "JustEat apologises for AI-generated slop after bizarre takeaway pictures found on Irish menus," October 2026. TechRadar, "Just Eat's AI put bacon and eggs in a glass of cola," October 20, 2026.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren