Quick answer: AI can suggest a name for food visible in a photograph, but a plausible dish name is not a full ingredient list, an allergen check or a measured portion. Compare the suggestion with what was actually cooked, the current product label or information from the food business. A barcode can identify a packaged product for a separate lookup; it does not verify its ingredients. Neither method establishes that food is suitable for someone with an allergy.
This guide explains how to review food identity, not how to calculate calories or choose a diet. It can be used with any photo-based suggestion. It does not claim that Smaklig UK currently offers photo recognition, barcode lookup, a nutrition database or a particular image-processing or privacy arrangement. Swedish Smaklig features are not UK product evidence.
AI-generated content. Always verify ingredients and allergen information yourself. The examples below are illustrative, not photos we tested in a UK kitchen or proof of any app's performance.
What does an image suggestion tell you?#
A food image model may suggest a broad category (“pasta”), a dish name (“spaghetti bolognese”) or a list of visible components (“pasta with red sauce”). These answers differ in detail. Even a correct dish name does not show that every ingredient was recognised. Shape, colour and texture are clues, while ingredients hidden inside bread, under sauce or in a sealed pack may be absent from the image.
Consider beans on toast. A photograph may make the beans and toast obvious, while any spread beneath the beans remains hidden. A takeaway curry may look recognisable, yet the recipe, cooking oil and kitchen handling are not visible. Two unlabelled white foods in bowls might be yoghurt, crème fraîche or something else. These are examples of missing evidence, not results from a test of Smaklig UK.
| Example | What the picture might support | What you still need to establish |
|---|---|---|
| Beans on toast | The visible beans and toast | Bread ingredients, any hidden spread and how much was eaten. |
| Takeaway curry with rice | Appearance of rice and sauce | Sauce ingredients, cooking additions and allergen handling by the business. |
| Unlabelled white food | Colour and texture | The actual food identity from its container or preparer. |
| Opaque ready-meal pack photographed from the front, with no readable ingredients label | Brand and front-of-pack design | Current ingredients and allergens. A clear photograph of the back label may show that information; read it directly. |
What did a controlled study actually measure?#
In a 2024 primary study of commercial nutrition apps, researchers submitted 22 food and drink images to seven selected apps with image-recognition features. The study compared the identification of food components and estimates of energy. Its images were taken under controlled lighting and background conditions, at a consistent angle and distance. The researchers reported difficulties with components of some mixed dishes; an eggs-on-toast example omitted butter in some app outputs.
Those are findings about the tested app versions and photographs, drawn from an Australian app sample. They do not measure Smaklig UK, today's version of any other app or a meal photographed in a busy kitchen. The authors warned that the controlled picture conditions and dietetics-trained researchers may not resemble everyday use. We therefore give no general AI accuracy percentage. A component-identification score also cannot establish accurate energy, portions, ingredients or allergy suitability.
Why is a portion a separate question?#
Naming “rice” does not tell you its weight, what was added during cooking, or how much remained on the plate. A single ordinary picture has no measurement of the amount eaten. A photo-based portion suggestion can be a prompt to check against a measured amount, pack or recipe, but should not be presented as a measured fact. The same applies if a model correctly names a whole dish while missing an oil or spread used in its preparation.
The research study evaluated energy estimates separately from component recognition. That distinction matters whenever an app gives a precise-looking nutrition result after a photo: the name, components, quantity and food-data record can each introduce uncertainty. A good-looking dish label cannot validate all four steps.
Is a barcode better evidence for a packaged product?#
It answers a different question. GS1's product-identification standard describes a GTIN as an identifier for a trade item. Reading a packaged product's barcode can lead a separate catalogue to a product record; an ordinary food photo suggests what its visible content resembles. A successful barcode read does not itself supply nutrition, ingredients or allergens, and a third-party record still has to match the current pack.
For a packaged item, check the physical product name, variant and label. GOV.UK's ingredients guidance describes ingredient lists and highlighted allergens. A picture of the pack's front with no readable ingredient text cannot settle the contents, while a clear image of the actual ingredient label may provide text you can read. Neither an AI guess nor a catalogue entry should replace the current pack when ingredients or allergies matter.
How can you review a photo suggestion safely?#
- State the task. Do you need a rough dish name, a packaged product's exact identity, an ingredient list or a measured quantity? The evidence needed changes with the task.
- Compare what is visible. If a suggestion says “yoghurt” but the bowl could contain another white food, ask the person who prepared it or check its container before treating the name as fact.
- Find concealed ingredients elsewhere. For home cooking, use the actual recipe and ingredients. For a pack, read its current label. For a restaurant or takeaway, ask the business about ingredients and handling.
- Keep quantity separate. A correct name does not reveal the weight eaten, added oil or leftovers. Mark an unmeasured amount as uncertain.
- Reject absence claims from silence. No visible nuts, milk or sesame does not show that they are absent. An image suggestion is not an allergen check.
GOV.UK's advice for eating out with a food allergy says restaurants and takeaways must provide information about the listed allergens and describes asking about kitchen handling. Follow the actual business's information and appropriate personal advice when an allergy is involved. A photo cannot establish cross-contact or whether a recipe changed.
What should an “AI accuracy” claim include?#
Ask which foods and app versions were tested, whether the photos resembled ordinary use, what counted as a correct answer and whether the score was for dish name, separate components, quantities or nutrition. A result from controlled images or another country's app catalogue is not a prediction for your takeaway or a different service. Until a current UK-relevant product evaluation states those details, leave claims about Smaklig UK's recognition rate open.
The useful next step is to match each question to its proper evidence: photo for a tentative visible description; barcode and current pack for packaged identity; recipe or food business for concealed ingredients; actual portion for quantity. Keep any unresolved part unresolved rather than accepting a confident-sounding answer.