AI Calorie Apps Underestimate Meals by 345 Calories
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- NIH researchers tested four photo-based AI calorie apps — MyFitnessPal, LoseIt!, CalAI, and Appediet — against 102 meals prepared in a controlled metabolic kitchen at the NIH Clinical Center with ingredients weighed to the nearest 0.1 gram.
- The four apps underestimated total calories by 250–345 calories per meal on average and fat by approximately 30 grams, with overall calorie and fat estimates running about one-third too low.
- MyFitnessPal and LoseIt! were more accurate when analyzing higher-calorie meals than lower-calorie ones, while all four apps produced more consistent estimates for carbohydrates than for other macronutrients.
- High-fat ketogenic meals appeared to trip up the apps most, with fats consistently underestimated; the researchers suggest the bias stems from the apps' difficulty evaluating fat-heavy dishes.
- Researchers followed up by testing more than 200 additional meals to identify which factors drive inaccuracy, finding preliminary evidence that keto-style meals cause the largest errors.
- Olivia Charles of NIDDK presented the findings at NUTRITION 2026 on July 25 in work led by Aaron Hengist; the results are preliminary and have not yet completed full peer review.
Why it matters: For the millions using photo-based apps to manage weight or metabolic health, a 250–345-calorie daily underestimate means the app is logging roughly a full meal's worth of food less than what was actually eaten — a gap that could quietly stall weight loss or skew clinical nutrition tracking.
