AI Food Photos: Check a Menu or Delivery App Photo

0

๐Ÿ”Ž Reverse Image Search

Image search techniques made simple: upload a photo or paste a URL and search 6 engines at once. 100% free, nothing stored.

๐Ÿ”’ Your image is never stored on our servers longer than 24 hours and is never shared.

A menu or delivery app photo can be AI-generated, AI-enhanced, or a real photo of a different dish. Check it in three places: look for impossible gloss, repeated garnish and wrong shadows; reverse search the image; and read the platform rules, because Deliveroo bans AI images outright while DoorDash labels its enhancements. No single check proves a photo is real.

If you want to check AI food photos, menu delivery app listings are where the risk sits. The picture is the only thing you see before you pay, and in October 2026 Just Eat had to apologise after its own AI tool put bacon and eggs in a glass of cola on a real menu. This guide shows you how to read a menu photo properly, what each delivery app actually allows, and what to do when the food that arrives looks nothing like the picture. Dish identification, working out what a dish in a photo actually is, stays with our reverse image search for food guide. This article is about something narrower: whether the photo itself can be trusted.

What is an AI food photo?

Answer capsule: an AI food photo can mean three different things, and they matter very differently when you order.

People use the phrase for three separate cases. A fully generated image is made from a text prompt: no plate, no kitchen and no dish ever existed. An AI-enhanced photo starts as a real photograph of a real dish, then an AI tool changes part of it: a new background, a tidier plate, a different angle, or garnish that was never there. A retouched photo is a real photograph with ordinary edits: brighter light, sharper detail, a cleaner crop.

Only the first case invents the dish. The second case is where most arguments live, because the dish is real but the picture has been improved past what the kitchen serves. The third case is normal food photography and has existed for decades. When you check a menu photo, your first job is to work out which of the three you are looking at, not to shout AI at anything that looks polished.

That distinction is also the law in the background. In the UK, a food picture is a claim about size, quantity, ingredients and appearance, and the advertising rules judge it like any other claim. In the EU, the AI Act requires AI-made images that look like real objects and could pass as genuine to be disclosed. The platform rules add a third layer, and they differ from app to app, which is why the policy table later in this guide matters.

What happened when Just Eat tried AI menu photos?

Answer capsule: in October 2026 Just Eat apologised, removed the images and paused its AI menu photo tool after the pictures it generated were ridiculed online.

Just Eat had been trialling a tool that generated a picture from a text description for menu items that had no photograph, so small restaurants could show a dish without paying for a photoshoot. The idea was reasonable. The results were not. The tool read dish names literally and invented food that nobody cooks.

The BBC reported the story on 21 October 2026. Its examples show how literal the tool was. A bacon and egg soda from a takeaway in Ireland appeared as a glass of cola with bacon and a fried egg floating in it. Chip muffins were shown as muffins with potato chips baked into them, served with brown sauce. Salt and pepper chips appeared as plain chips covered in salt and pepper, when the real dish usually includes onions, garlic, chillies and other ingredients, often with curry sauce. A cheese roll was drawn as sushi rolls containing cheese. Garlic bread listed as plain was shaped like an aeroplane.

Just Eat apologised. A spokesperson told the BBC: “We know these images fell short of the standards our customers rightly expect and have removed them. We have also paused the tool that generated them while we work with the restaurant to put this right.” The BBC also reported that the pictures were not labelled as AI-generated, so customers had no easy way to tell they were looking at a generated image rather than a photograph.

The episode is useful for two reasons. First, it shows the failure mode of generated food images: they go wrong on meaning, not just on pixels. The tool did not draw a bad burger. It drew the wrong dish entirely, because it read the words instead of knowing the food. Second, it shows why labels matter. A customer who sees a label can decide for themselves. A customer who sees no label assumes the photo is the dish.

For the full BBC report, see the BBC story on the Just Eat apology: Just Eat apologises for AI food mistakes after social media ridicule.

How can you tell if a menu photo is AI-generated?

Answer capsule: look for a cluster of small impossibilities, not one dramatic flaw, because a single strange detail can also come from retouching or compression.

Start with the food itself. Real food is messy in specific ways. Sauce drips, herbs fall at random angles, grill marks vary, and steam softens edges. Generated food is often too even: every seed identical, every herb leaf placed at the same angle, every grill mark the same depth. Count the toppings. If ten identical basil leaves sit in a neat row, or every sesame seed on a bun looks like a copy of the next one, that is a clue, not proof. Real kitchens plate by hand, and hands are inconsistent.

Next, look at the gloss. Generated food often shines in a way real food does not. A burger bun can look varnished, cheese can look wet and plastic at the same time, and a sauce can glow with a highlight that has no light source behind it. This is the tell most people notice first, and it is also the weakest on its own, because a real photo with heavy retouching can shine too. Use gloss as a reason to look closer, never as a verdict.

Then check the small objects around the plate. Fork tines that bend or merge, chopsticks that end in mid-air, a knife handle that changes shape, a glass rim that is not a circle, lettering on a menu board that turns into nonsense shapes: these are the places generators still struggle. Zoom in on the highest quality version of the image you can open. A small thumbnail in the app can turn ordinary compression into apparent defects, and it can also hide real ones.

Shadows and reflections come next. Trace one light direction through the whole scene. The highlight on the glass, the shadow under the plate and the reflection in the cutlery should all agree. A plate that floats above the table with no contact shadow, a shadow that falls the wrong way, or a reflection that omits an object sitting right next to it are all worth noting. Again, one odd shadow proves nothing. Three independent oddities in one image start to mean something.

Finally, read the background. Signs, labels, menus and logos in the background often carry malformed characters in generated images. Chairs, tiles, bottles and tableware can fuse together or change shape at the edge of the frame. Depth should make sense: a plate, a table and a wall should share one believable perspective.

A study published in 2024 in Food Quality and Preference, by Califano and Spence, tested how well people could tell real and generated food images apart. Across the food items tested, correct identification ranged from 62 percent to 91 percent, with category averages of 78 percent for unprocessed food, 78 percent for processed food and 83 percent for ultra-processed food. That spread is the honest picture: some generated food photos are easy to spot, and some are not. Visual checks are screening, not certification. Our guide to how to tell if an image is AI-generated covers the same visual method for any kind of picture, and the same limits apply here.

What do Deliveroo, DoorDash, Uber Eats and Just Eat actually allow?

Answer capsule: Deliveroo bans AI-generated images outright, DoorDash labels its own AI enhancements, Uber Eats demands a photo of the real item, and Just Eat paused its own AI tool in October 2026.

Platform rules change fast in this area, so every entry below was re-checked on 14 November 2026 against the platform’s own pages. If you are reading this later, check the live page before you rely on it.

PlatformWhat the rule saysWhat it means for you
DeliverooAI-generated images are not permitted. Partners must only upload authentic images of the food or products they sell.Any AI-generated dish photo on Deliveroo breaks the platform’s own rule. Report it through the app.
DoorDashMenu photos must show the actual item served, not an idealised or generic version. DoorDash labels photos it has AI-enhanced.An “AI-enhanced” label means DoorDash changed the image beyond basic lighting or clarity. The food itself should be unchanged.
Uber EatsItem photos must accurately represent a single item from the Uber Eats menu. No people except hands, no extra items in frame.Uber Eats has no AI clause in its photo rules, but the dish in the photo must be the dish that arrives.
Just EatJust Eat trialled AI-generated images for items without photos, then apologised and paused the tool in October 2026.AI images were still visible on some menus when the BBC reported the story. Check the dish, not the label.

Deliveroo’s rule is the clearest of the four. Its Help Centre page on uploading photos in Menu Manager, updated 2 September 2026, says in plain words: “AI-generated images are not permitted on Deliveroo. Please only upload authentic images of the food or products you sell.” The same page sets the technical rules: JPEG files, landscape, at least 1200 by 675 pixels, up to 8 MB, cropped to a square in the app. Deliveroo also offers to enhance real photos that have a cluttered background, blur or poor lighting, which is a different thing from generating a dish. Read the full rule on Deliveroo’s Help Centre photo page.

DoorDash sits in the middle. Its consumer help page defines the “AI-enhanced” label like this: if you see “AI-enhanced” on a photo, DoorDash has used AI to improve the image in ways that go beyond basic lighting or clarity adjustments. Its examples include re-plating the dish, changing or adding a background, filling in missing parts of the image, and adjusting the perspective. DoorDash also says it is focused on improving photos without altering the actual food, ingredients or portion size, so the dish you see is representative of what the merchant offers. That is the company’s stated policy and intent, not independent verification of every listing. On the merchant side, DoorDash announced three AI photo tools in May 2026: AI Retouch, which cleans up backgrounds and lighting while keeping the food as it is; AI Replate, which re-stages the dish on different dinnerware; and a style-matching tool. Its merchant guidance says images must reflect the actual item served, not an idealised or generic version, and its virtual brand rules add that evidently AI-generated images are not permitted. Item photos should be landscape 16 by 9 and at least 1400 by 800 pixels.

Uber Eats has no AI clause in its merchant photo guidelines, but the accuracy rule does the same work. Item photos must accurately represent a single item from the Uber Eats menu, be framed in the centre, sit between 5:4 and 6:4 aspect ratio, and show no people except hands. Files can be JPG, PNG or GIF up to 10 MB. Uber has also said it uses AI to detect and enhance low-quality food images, improving lighting, resolution, framing and plating.

Just Eat is the cautionary tale. Its own tool generated the images, labelled or not depending on the report, and the company paused the rollout while it worked on the tool. If you still see a suspicious image on Just Eat, the dish check in the next section is the useful test, because a label may or may not be there.

One more number is worth knowing, because it explains why this matters to restaurants. DoorDash says menu items with photos see up to 44 percent more monthly sales than items without, according to its Merchant Learning Center. That is a strong reason for a restaurant to add a photo, and a strong reason for a platform to fill the gap with AI when a restaurant has none.

How do I check a menu photo before I order?

Answer capsule: run the eight-step menu photo check set, and treat the photo as suspicious only when several checks fail together.

This is the check set to use on any menu or delivery app photo that matters to you, whether it is an allergy, a religious requirement, a portion dispute or simply a dish you are paying good money for. It takes about two minutes once you know it.

  1. Read the words beside the photo. Compare the dish name, the listed ingredients, the stated portion and the serving style with what the picture shows. A two-topping pizza showing five toppings, a soup described as clear but pictured as creamy, or a family portion shown in a tiny bowl all deserve a question before you order.
  2. Zoom in on the small things. Open the largest version of the image you can. Check fork tines, glass rims, plate edges, garnish and any text in the background. Note every anomaly, and only count it when you can describe exactly what is wrong.
  3. Trace the light. Follow one light direction through the highlights, shadows and reflections. One wrong shadow is a note. Two or three that disagree with each other is a pattern.
  4. Check the restaurant’s other photos. Open other dishes from the same restaurant and, where the app shows them, customer photos of the same dish. Consistent dishware, counter surfaces and portion sizes support a real house photo, though they do not prove it. A sharp difference can mean a new photographer, an old photo, a stock image or AI editing.
  5. Reverse search the image. Search the full image and a crop of the dish itself. Look for an earlier appearance, a stock photo page, another restaurant using the same picture, or a packaged product. A photo that appears on five different restaurant menus was not taken in any of their kitchens. Our main how to reverse image search guide walks through the tools step by step.
  6. Check for a label or provenance signal. Look for an “AI-enhanced” label on DoorDash, an AI-generated label where a platform shows one, or Content Credentials data if you have the original file. A positive signal tells you what that system recorded. No signal tells you nothing at all, because most images carry no provenance. Our Content Credentials check guide explains how to read the file data when it survives.
  7. Run a detector, and distrust a clean result. An AI image detector can support a suspicion, but a “real” result never clears a photo. Detectors are trained on known generators, they struggle with edited real photos, and app compression strips the signals they look for. Treat the detector as one vote, not a verdict.
  8. Ask before you accuse. If the photo affects a purchase and it is still unclear, message the restaurant: “Is this a photograph of the dish you currently serve, or a generated or enhanced illustration?” Save the listing, the date and the reply. A straight answer settles it. No answer is also an answer.

Worked example: a real burger and a Just Eat AI image, checked side by side

Answer capsule: the real burger traces back to a dated original with camera data, the Just Eat AI image traces back to the BBC story that documented it, and the detector scored both almost the same, so it could not tell them apart.

To show what the check set actually produces, I ran two food photos through the same steps on 14 November 2026. The first is a real cheeseburger photograph from Wikimedia Commons, released under a CC0 licence by the photographer Kurtkaiser, taken on 22 May 2021. The second is the BBC’s published image of two Just Eat AI-generated menu pictures from October 2026: chip muffins on one side and a glass of fizzy drink with bacon and egg floating in it on the other.

The real burger first. Its file is a 4608 by 3456 pixel JPEG with camera data intact, including the date it was taken, 22 May 2021 at 20:40. That alone does not prove a photo is real, but it is the kind of ordinary history a genuine photograph carries. Google Lens found 11 exact matches, and the top two were the same photo on Wikimedia Commons, under its own file name Cheeseburger 2.jpg, at the same 4608 by 3456 pixels. Lens described it as a loaded gourmet cheeseburger on a wooden board. TinEye found nothing at all: 0 matches across the 86.7 billion images it has indexed, even after a second upload gave the same result. That is a useful warning in itself, because TinEye had simply not crawled the pages where this photo appears. The Hive detector, which now runs as Deploy Detect, scored it 0.2 percent likely to be AI-generated and 7 percent likely to be a deepfake, with the verdict that it is not likely to contain AI-generated or deepfake content, and it found no Content Credentials.

The Just Eat image is a different kind of file. It is a 480 by 270 pixel image taken from the BBC’s article page, and it carries no camera data, no C2PA record and no embedded description at all, which is normal for a web image but means the file itself tells you nothing. Google Lens found 72 exact matches, and the top result was the BBC story itself, Just Eat apologises for AI food mistakes after social media ridicule, dated 21 October 2026, followed by the same story republished on Instagram, AOL and other outlets. Lens identified the picture as showing bizarre AI-generated food pictures, chip muffins with gravy and bacon and eggs floating inside fizzy drink, that prompted an apology from Just Eat. TinEye found 4 results, all of them the BBC, first indexed on 21 October 2026. The detector, though, missed it completely: Deploy Detect scored the known AI-generated image 0.8 percent likely to be AI-generated and 0 percent likely to be a deepfake, and gave it the same not likely verdict as the real burger. A detector that cannot tell a documented AI image from a real photograph is not a test you can rely on alone.

Three lessons fall out of the pair. First, a reverse search is most useful when it finds something: the Just Eat image traced straight back to the BBC story that documented it, while the real burger traced back to its own Commons page on Lens but nowhere on TinEye, which shows that coverage differs between tools. Second, the detector failed outright: it scored the documented AI image 0.8 percent likely to be AI-generated against 0.2 percent for the real burger, and gave both the same not likely verdict, so a clean detector result is not a clearance. Third, the strongest evidence in the whole exercise was not a tool result at all. It was the BBC’s reporting, which documented that the Just Eat images were generated, named the dishes, and recorded the apology. When a photo comes with a public record like that, read the record.

What should I do when the dish arrives looking nothing like the photo?

Answer capsule: photograph what arrived, keep the listing, ask the platform for a refund first, and escalate only if the platform will not act.

Start before you eat. Photograph the dish as it arrived, next to the menu photo on your phone screen if you can, so the difference is visible in one frame. Screenshot the listing, including the dish name, the description and the price, before it changes. Note the time the order arrived and keep the receipt.

Then contact the platform, not the restaurant first, because the platform holds the money and the refund route. On Deliveroo, Just Eat and Uber Eats, use the in-app help for that order and choose the option for an item that does not match its description. Describe the difference in plain words: what the photo showed, what arrived, and what is missing or different. Ask for a refund or a redelivery. Keep the message factual and keep a copy.

If the platform refuses or offers only a partial refund, your next step depends on where you are. In the UK, a misleading food picture is an advertising matter, and you can report it to the Advertising Standards Authority, which handles misleading ads, or to Trading Standards through Citizens Advice if the listing itself misled you. The Consumer Rights Act says goods must be as described, which includes the picture that sold them to you. In the US, the route is the Federal Trade Commission’s ReportFraud site, plus your state attorney general’s consumer protection office. A card chargeback through your bank is also available if you paid by card and the item was materially different from what was described, but try the platform first, because a chargeback takes longer and the platform may close your account if you use it as a first step.

One caution: do not publicly accuse a restaurant of using AI from one photo. A mismatch can come from a stock photo, an old menu image, a different chef on the night, or a portion that shrank since the photo was taken. Accuse the picture of misleading you, which you can show, not the kitchen of generating it, which you cannot prove from one image. If the restaurant confirms the photo is generated or enhanced, that confirmation is worth keeping, because it is the one piece of evidence no tool can give you.

What mistakes do people make when checking food photos?

Answer capsule: the four common mistakes are treating one anomaly as proof, trusting a detector’s clean result, reading missing metadata as AI, and confusing a reused stock photo with a generated one.

The first mistake is treating “too perfect” as proof. A polished image is not evidence by itself. Professional food photographers light, style and retouch real dishes until they look unreal, and a good phone photo in bright light can look generated to an untrained eye. Use perfection only to decide where to inspect, then look for a cluster of inconsistencies and corroborate them with the listing and the source history.

The second mistake is trusting a detector when it says “real”. A failed check cannot rule out generation, enhancement or signal loss during upload. Detectors are trained on specific generators and file types, and a delivery app compresses every image it serves, which strips the fine detail detectors look for. A detector that says “likely AI” is useful supporting evidence. A detector that says “likely real” is almost no evidence at all.

The third mistake is assuming missing metadata means AI. Most images lose their embedded data the moment they are downloaded, resized, screenshotted or uploaded to an app. A menu photo with no metadata is normal, not suspicious. Only an original file with intact data can tell you anything, and even then the absence of a provenance record proves nothing, because most real photos have none either.

The fourth mistake is confusing a reused stock photo with a generated image. A restaurant that uses a stock photo of a burger is misleading you about its own dish, but the photo itself is real. Reverse search is the tool that separates the two cases: a stock page or another restaurant’s menu explains the mismatch without any AI involved. If you find the same photo on a stock site, the problem is reuse, and the complaint is the same, but the fix is different: ask for a photo of the actual dish.

Product photos have the same problem in a different aisle. Our AI-generated product photo check covers the same tells for shopping listings, where the stakes are a purchase rather than a meal.

FAQs

Can a restaurant use an AI-generated photo on Deliveroo?

No. Deliveroo’s Help Centre page, updated 2 September 2026, says AI-generated images are not permitted and asks partners to only upload authentic images of the food or products they sell. Deliveroo does offer to enhance real photos that have a cluttered background, blur or poor lighting, but that is an edit to a real photo, not a generated dish.

What does AI-enhanced mean on DoorDash?

On DoorDash, an AI-enhanced label means DoorDash used AI to improve the image in ways that go beyond basic lighting or clarity. That can include re-plating the dish, changing or adding a background, filling in missing parts, or adjusting the perspective. DoorDash says it does this without altering the actual food, ingredients or portion size.

Can I tell from one look whether a food photo is AI?

Usually not. A study in Food Quality and Preference found people identified generated food images correctly between 62 and 91 percent of the time, depending on the dish, which means some are easy to spot and some are not. Use the full check set: visual tells, the listing, a reverse search, and the platform rules together.

Does a clean reverse image search mean the photo is real?

No. A clean reverse search only means the image has not been indexed elsewhere yet, or that the search engine has not matched it. A generated image made this week may have no earlier appearance at all. A clean result is the absence of evidence, not evidence of a real photo, so combine it with the visual and listing checks.

What should I do if my delivery looks nothing like the menu photo?

Photograph what arrived and screenshot the listing before it changes, then contact the platform through the in-app help for that order and ask for a refund or redelivery. Describe the difference plainly. In the UK you can also report a misleading listing to the Advertising Standards Authority or Trading Standards through Citizens Advice.

Are AI food photos illegal in the UK?

There is no law that bans AI food photos, but a misleading picture breaks the advertising rules, because a food photo is a claim about size, quantity, ingredients and appearance. In the EU, the AI Act requires AI-made images that look real and could pass as genuine to be disclosed. The platform rules add their own bans and labels.

Ali Butt is the lead writer and SEO specialist at Image Search Techniques, where he covers reverse image search, visual search, and online image verification. With over four years of experience in SEO and content, he writes practical, hands-on guides that help readers trace a photo to its source, identify people, products, plants and brands from a picture, protect their own images online, and spot catfish, fakes and AI-generated images across Google, Yandex, Bing, TinEye and more.

Leave A Reply