The Years of Magical Thinking: Just Eat, AI menu images and liability
We're three years into the wave of genAI-enabled transformation, and people still don’t seem to understand that the machine is not magic or godlike. It is an unfeeling and unthinking statistical system. The limits of AI systems need to be acknowledged up front - there will be errors, and you must put in place measures to assure the content and mitigate risk.
Just Eat - a food delivery platform in the UK - recently provided us with a very public reminder.
What happened?
Just Eat used AI to generate images for menu items where restaurants hadn’t supplied photography. The AI's interpretations of different dishes were... questionable in some instances.
Tiger prawns were interpreted as small tigers (I suspect that the system didn't distinguish between prawns and shrimp, and made the lexical jump from 'shrimp' to 'small'). A jacket potato was considered to be more jacket than potato. A bacon and egg soda bread sandwich, listed as the 'Breakfast Soda', was depicted as bacon and eggs, served sparkling, in a glass.
A misspelt 'plain' garlic bread became a bread-shaped Boeing.
Just Eat has reportedly apologised, begun removing the images and paused the rollout of the AI-images.
The images carried an 'AI generated' label. Just Eat claims that each image was checked before publication, but that is difficult to believe.
We cannot blame the AI system. The system has never had breakfast, and unless it's specifically given adequate context, will likely not understand the regional or cuisine-specific idioms that abound in food culture.
The blame truly lies on organisations who don't understand the limits of the technologies they're using, or refuse to manage the risks.
The risk
Beyond the obvious incompetence (and high comedy), I think there is a significant compliance risk here.
A menu image has a specific role: it tells a customer what they are buying. In Australia, the ACCC is clear that businesses must not mislead consumers, and that includes images and descriptions of what is offered. Intent is not required.
Nothing about this is AI-specific. Treasury’s Review of AI and the Australian Consumer Law found the ACL is technology-agnostic and broadly capable of addressing AI-enabled conduct. Similarly, the Department of Industry’s guidance says the same: businesses must ensure their representations are truthful, including where AI was involved.
The ACCC has not prohibited AI imagery in retail contexts, but you own the outcome - if your AI image is misleading, you could breach s18 (misleading or deceptive conduct) or s29 (false or misleading representations) of the Australian Consumer Law. In particular, s29 has maximum penalties of up to $100 million or 30% of annual turnover for corporations (s224 ACL).
The tigers are also not the real risk. Nobody is being misled by images of live tiger cub stew. The risk is the plausible image: a generous-looking portion, or a dish garnished in a way the kitchen never would. Absurd errors get caught, and plausible ones won't...or at least, not immediately. Just Eat did label the images, but a label won’t necessarily save you either, because the ACCC looks at the overall impression, and disclaimers may not remove risk if that impression remains misleading.
Labelling: worth mandating, but not a cure
There’s a good argument that we should be mandating labelling of AI images or enhanced images, as NSW has recently done with real estate imagery. Under the Residential Tenancies (Protection of Personal Information) Amendment Bill 2025 passed in September 2026, landlords and agents will need to disclose when rental images have been digitally altered or AI-generated to hide faults, such as fake furniture that conceals structural damage.
But labelling is a floor, not a solution. Just Eat labelled its images and still published flight-capable garlic bread. A label tells people how an image was made - but it won't make it accurate.
There's a reasonable argument that perhaps a 'truth in advertising' requirement may be required. This is hardly a revolutionary proposal - for example, Japan has had similar laws in place since 1962. Under the Act Against Unjustifiable Premiums and Misleading Representations companies are required not to exaggerate what's inside the packaging, especially when it comes to size, appearance, or quality. But when the exaggeration can be automated in moments, intentionally or otherwise, regulators may consider that there need to be more specific requirements than a general obligation to not be misleading.
So what?
AI solutions are here, and they're not going away - the temptation for fast and cheap outcomes is too beguiling.
But, if you're going to use them, you need to assess the solution and identify the limitations and risk early, and build in mitigations to address adverse outcomes.
An effective assessment would likely have identified the risks of inaccurate photos, and suggested requiring human review. Assessors would have then identified that adding an internal human layer would not be an effective fix - Just Eat has over 100,000 restaurant partners in the UK alone, and perhaps 30 times as many menu items on the platform. Because of the scale, an internal human review mechanism might quickly get overwhelmed. But there are ways around that too - Just Eat could have broken up the work and engaged with restaurant partners to require them to review their own menu images and flag inaccurate ones, or prompt them to replace images with their own photography. A stage gate could have been added to prevent the images from going live until each restaurant had confirmed that they were satisfied.
Situations like this are eminently avoidable - if you can start from the position that AI is a technology tool, with strengths and weaknesses, and not a magic wand.
Sabirus Advisory helps organisations govern AI technologies including agents and automated systems. Get in touch if you'd like to talk through where your organisation stands.
