Calorie-counting apps miss hundreds of calories per meal

Researchers ran 102 lab-prepared dishes—each ingredient measured to exactly 0.1 g—through the four leading apps: MyFitnessPal, Lose It!, CalAI and Appediet. On average they under-reported by 250 to 345 calories per dish and omitted roughly 30 g of fat. Carbohydrates were identified reasonably well, but as soon as the meal shifts to low-carb, high-fat territory the algorithms fall apart completely.

The pipeline looks tidy on paper: food identification, portion-size estimation, cross-reference with a database. In practice each step is a minefield, and errors compound. Fat packs about twice the calories of a carbohydrate or protein (9 kcal per gram versus 4), so missing a spoonful of oil or a slice of cheese erases hundreds of calories from the log. The AI doesn’t “see” hidden fat in sauces, frying, or melted cheese; it merely guesses from appearance, which sounds exactly like something an Israeli would say.

The authors of the preprint aren’t selling a magic bullet. The takeaway is simple: automatic counting is a starting point, not a safety net. Anyone who counts seriously—athletes, people with medical conditions, or anyone on a tight calorie deficit—needs to manually adjust ingredients and weights. Otherwise you end up in a phantom deficit while actually eating maintenance. A waste of time.

The technology is improving, but the gap between “looks like 200 g chicken breast” and an actual 200 g portion is still too wide. Until models learn to spot hidden fat and gauge volume without a reference, these apps remain a convenient aid, not an authority. Snap a photo, log it, but double-check. Go eat something.