Will AI-generated photos get your menu rejected?

DoorDash publishes eleven rejection reasons. The one that matters here is "photo is not representative of the item", which covers images that appear artificial, AI-generated in a way that does not show the actual dish, or heavily modified. Alongside it sit the ones that catch most people: wrong size or resolution, too zoomed in or out, poor lighting, problematic background, overlays or watermarks, people in frame, out of focus, copyright, and duplicates.

The rule underneath all of them is the same. The photo has to show what the customer will actually receive. That is a representation rule, not a technology rule.

Uber Eats requires the item centered with nothing running off frame and no more than one item per photo. Grubhub requires food photos only, and checks copyright by looking for the image in a reverse image search, which is worth knowing if you are tempted to pull something off the internet.

Because rejecting generated food and offering enhancement are not in conflict, and DoorDash drew the line itself.

In April 2025 DoorDash announced a set of AI photo features for merchants. Read what they say the tools do, in their own words.

That phrase appears twice, in two different product descriptions: without altering the food itself. The platform that will reject your generated photo is the same platform telling you exactly what it considers acceptable.

So the question to ask about any photo tool is not whether it uses AI. It is what the tool is allowed to touch.

The awkward cases are the vessel and the portion, and they are where well-meaning tools drift. A tool that restages your food onto a nicer plate has changed something the customer will notice when the box arrives. So has one that quietly adds a garnish because generated food usually looks better with one.

This is the test I use on our own output: if the dish that arrives would make a customer feel misled, the result gets thrown out, no matter how good it looks.

Here is that test failing, on our own pipeline.

Two separate risks, and the second one is worse than the first.

In July 2026 a San Francisco cafe called Grind & Unwind put AI-generated food images on a window board. The storefront was defaced with graffiti. The owners removed the board, repainted at a cost of around $700, and replaced it with LED lights and a display of their actual pastries. Commenters online called it false advertising and picked the images apart in detail, down to bread that looked like reptile skin.

The owner's own reaction is the part worth sitting with: "I want people to come to our restaurant, but, like, why is this such a big deal that it influences people on such an intense level?"

It is a fair question, and the answer is that a menu photo is a promise about what arrives. People react to a broken promise about food more strongly than the person making it usually expects.

The honest answer is that the research points in two directions, and anyone selling you something who tells you otherwise is picking the half that suits them.

Both can be true. An image can be more attractive in isolation and less convincing when a person is deciding whether to trust it, and separate work on the uncanny valley in food imagery points the same way: the closer a synthetic image gets to real, the more a small wrong detail costs it.

Which of the two you get is decided by context, and a menu is the worst possible context for the first effect. The customer is not admiring the image. They are about to pay for the thing in it, and then compare.

Photograph the actual dish, then fix the photograph rather than the food.

A phone photo taken near a window, with the dish framed with margin on all four sides, gives you something real to work from. From there, lighting, background, colour and crop are all fair game, and they are what separates a usable menu photo from an unusable one. The dish stays exactly what you served.

If you want the platform sizes handled at the same time, that is a separate problem and it has its own guide.