Apps that can estimate the calories of a meal from a simple photo promise rapid dietary tracking. However, an evaluation conducted on four popular services shows that they often underestimate the energy and fat content on the plate.
The researchers tested MyFitnessPal, Lose It!, Cal AI, and Appediet with 102 meals prepared in a metabolic kitchen at the National Institutes of Health. Each ingredient had been weighed to within 0.1 g, providing a much more accurate reference than an ordinary visual estimate.
Illustration: Pixabay
For each meal, a standardized photograph was sent to the four apps. Their results were then compared with the actual nutritional values. On average, the estimates showed between 250 and 345 fewer calories per meal, a reduction of about one-third.
The errors particularly involved fats, which were underestimated by about 30 g per meal. Carbohydrates were generally better recognized. Early results also indicate that dishes low in carbohydrates and high in fat, such as certain ketogenic meals, cause more inconsistencies.
A photograph does not always reveal the amount of oil, butter, sauce, or ingredients hidden in a dish. Artificial intelligence also has to estimate volumes from a two-dimensional image, without precisely knowing the recipe or the density of the foods.
These discrepancies can skew a food diary used to lose weight, adjust a treatment, or discuss with a healthcare professional. The authors therefore recommend checking portions and manually completing the information when nutritional accuracy is important.
The study remains preliminary. It was presented at the NUTRITION 2026 congress and compares four apps under controlled conditions, without directly measuring their effects on users' eating habits.
The researchers continued the analysis with more than 200 additional meals to identify the types of dishes that are hardest to interpret. This work could help developers improve portion estimation, fat recognition, and the integration of user-provided information.