How accurate are AI calorie counting apps?

Three places. Recognition: is that chicken or turkey, wholegrain or white? It works well for visible components and badly for hidden ones: oil in the pan, sugar in the dressing, butter under the vegetables. Portion: a photo gives area, not height or density; 150 g of rice and 220 g of rice look almost the same in a deep bowl. Reference: which nutrition table the app uses for "chicken breast" moves the answer by 20–30 kcal per 100 g on its own.

Kairo narrows the third source by cross-referencing USDA FoodData Central and Germany's DGE values, and by pulling label values for branded products instead of estimating them. The second source, portion, stays the honest uncertainty. That is why correcting a portion before saving is a single tap in Kairo rather than a menu.

On 16 June 2026 we ran 16 meals spanning German, US, Japanese and Italian food through the production analysis path: the exact code that processes your photo, not a test mode. Each result was scored deterministically per 100 g against its reference, with a programmatic check that every source the analysis cited really exists.

Branded items against Open Food Facts: Kinder Joy 0.0%, Nutella 0.2%, Skyr 1.5%, Kölln oats 3.4%, Pocari Sweat 4.4%, Chobani 11.7%; mean ~3.5%, median ~2.5%. Restaurant against published chain figures: McDonald's cheeseburger 4.2%. Whole foods against USDA: apple 9.9%, chicken breast 10.1%, banana 12.4%. All 16 passed the ±20% gate; weighted-macro accuracy across the set was ~96%. It is a small internal set. A different set would shift the averages, and we say so.

Manual logging is only as accurate as the portion you type. Decades of dietary research show that people under-report what they eat. Not because they type badly, but because they forget the oil, the second helping and the snacks. A photo does not forget the second helping if you photograph it.

With a scale and a label you beat any photo app; that is the reference we score against. Without a scale (which is to say, in real life), the photo estimate is often closer to the truth than the portion you recall into a database search. The gain is less about any single number and more about doing it every day.

Shoot from a high angle so height is visible, and do not stack things. Say what is hidden, in text or by voice: "two tablespoons of olive oil", "oat milk", "half portion". Use the barcode instead of the photo for anything packaged. The label beats any estimate. And correct the portion when you know it; Kairo's model learns from your corrections how you actually eat.

Over weeks, something else matters more: Kairo derives your real energy expenditure from your weight trend. If the estimates run systematically 8% low, the learned expenditure corrects the target. The scale is the final arbiter, not the photo.

For fat loss and maintenance, yes. A 2,000 kcal target with 8% noise is about 160 kcal of uncertainty, less than one forgotten snack. For muscle gain with a protein target, also yes, because protein sources (meat, dairy, legumes) are easy to recognise. For medical requirements, such as carbohydrate counting on insulin, the decision belongs to you and your clinician; there, a scale is mandatory and no app is a replacement.

Our accuracy page also lists what the benchmark does not show, for instance, that we tested no other apps. If you want the checkable facts about how other apps handle accuracy, read Which AI food scanner is the most accurate?.