Taking a photo of a meal may feel easier than weighing ingredients or searching through a food database. But a preliminary study suggests that convenience may come with a sizable margin of error.
Researchers at the National Institutes of Health tested photo-based features in four calorie-tracking apps and found that all four underestimated the energy content of meals by roughly one-third. Fat was also consistently underestimated.
The findings were presented at NUTRITION 2026, the annual meeting of the American Society for Nutrition, and have not yet undergone the full peer review required for publication in a scientific journal. The study evaluated app estimates, not whether using the apps affected weight loss, eating habits or health.
The researchers tested MyFitnessPal, Lose It!, Cal AI and Appediet using photographs of 102 meals prepared for a controlled feeding study at the NIH Clinical Center.
Because the meals were prepared in a metabolic kitchen, researchers had detailed information about their ingredients and nutrient content. That gave them a more precise reference point than they would have had with meals prepared or reported by participants.
“By using meals prepared in a tightly controlled metabolic kitchen, we were able to compare the apps’ estimates against a precise reference,” said Aaron Hengist, a postdoctoral visiting fellow with the National Institute of Diabetes and Digestive and Kidney Diseases.
The meals averaged 918 calories, 61 grams of fat, 60 grams of carbohydrates and 36 grams of protein.
Compared with the known nutrient content, the apps underestimated calories by an average of:
327 calories for MyFitnessPal
333 calories for Lose It!
345 calories for Cal AI
252 calories for Appediet
All four also underestimated fat by about 30 grams per meal.
The apps were somewhat more consistent when estimating carbohydrates than when estimating calories, fat or protein. MyFitnessPal and Lose It! also tended to come closer on higher-calorie meals than on lower-calorie meals, though their estimates still showed substantial variation.
The size of the error differed considerably from meal to meal. That means an app might come reasonably close for one plate and miss the mark by several hundred calories for another.
“People using a photo-based tracking app without adjusting the portions or entering the amounts of food should take the results with a grain of salt,” Hengist said. “These apps tend to underestimate calories, especially from fats, so what they actually ate is likely higher than what the app shows.”
One likely challenge is that a photograph cannot always reveal what is in a meal.
Cooking oil, butter, creamy sauces, salad dressings and other sources of fat may be difficult to identify or measure from appearance alone. Two dishes that look similar in a photo can contain very different amounts of energy depending on how they were prepared.
Portion size can also be difficult to judge without a clear reference for scale. A photo may show what food is present, but not necessarily how much is on the plate or how ingredients are distributed throughout a dish.
The study does not show that the apps are useless. Automatic photo estimates may still help users create a rough food record or reduce the time required to log meals. But the findings suggest that users should not assume the first number generated by an app is exact.
Accuracy may improve when people review the foods identified, adjust portion sizes and manually enter ingredients that are difficult to see. The study did not test whether those additional steps corrected the errors.
It also did not evaluate the apps’ traditional manual logging tools. The findings apply specifically to the photo-based AI features tested, not to every way someone might use MyFitnessPal, Lose It! or another tracking platform.
Another limitation is that the meals were relatively large and high in fat. The average meal contained more than 900 calories, so the results may not apply in the same way to smaller or simpler foods such as a piece of fruit, a bowl of cereal or a plain sandwich.
The abstract also does not identify the precise software versions, subscription levels or device settings used. That matters because AI-powered tools can change quickly as companies update their image-recognition systems and nutrition databases.
The study’s central finding is still useful: A meal photograph alone may not provide enough information for an app to accurately calculate calories and nutrients.
For people who choose to track their food, photo-based tools may be most helpful as a convenient first estimate. They are less suited to serving as an exact accounting of what was eaten, particularly when a meal includes ingredients that are hidden or difficult to judge visually.
The research was supported by the Intramural Research Program of the National Institute of Diabetes and Digestive and Kidney Diseases within the National Institutes of Health.