Four Apps, One Direction of Error
Researchers at the National Institutes of Health photographed 102 meals whose exact composition was already known, ran the photos through four popular AI calorie tracking apps, and found that every app came back low.
The apps underestimated the calories in a meal by roughly 250 to 345 calories on average, according to research presented at NUTRITION 2026, the American Society for Nutrition's annual meeting. Fat was underestimated by about 30 grams per meal. Across the board, the shortfall came to about one third of the actual calorie and fat content.
The direction of the error is what makes this worth a reader's attention. A tool that is randomly wrong averages out over a week. A tool that is consistently wrong in one direction does not. Someone photographing three meals a day could be under by several hundred calories daily without any indication from the app that anything is off.
The apps tested were MyFitnessPal, Lose It!, Cal AI, and Appediet. Photo-based estimation has become a default feature in mainstream nutrition apps and is increasingly built into wearables, which means this is a measurement question affecting a very large number of people.
Why the Test Design Matters
The comparison here was unusually clean, and that is the study's main strength.
The meals came from a metabolic kitchen at the NIH Clinical Center, a research facility where every ingredient is weighed to the nearest tenth of a gram. The photographs were standardized. That means the researchers were not estimating the "true" nutrition of the meals. They knew it. The apps were being measured against a known answer rather than against another estimate.
The meals were drawn from a larger diet study examining how the body processes nutrients on a low-carbohydrate ketogenic diet compared with a standard diet. After the initial analysis of 102 meals, the researchers examined more than 200 additional meals to understand what drove the errors.
Aaron Hengist, a postdoctoral visiting fellow with the intramural program at the National Institute of Diabetes and Digestive and Kidney Diseases, described the practical upshot for users who snap a photo and accept whatever number appears. "These apps tend to underestimate calories, especially from fats," Hengist said, adding that what a person actually ate is likely higher than what the app displays. The findings were presented by Olivia Charles, a postbaccalaureate intramural research training fellow at the same institute.
What Is Established and What Is Not
This is preliminary work, and that framing belongs alongside the finding rather than after it.
The results were presented as a conference abstract. Abstracts submitted to NUTRITION 2026 were reviewed and selected by a committee of experts, but they have generally not completed the full peer review process required for publication in a scientific journal. The American Society for Nutrition states this directly in its own materials. Until the analysis appears in a peer-reviewed journal with full methods, the specific numbers should be treated as provisional.
There are also limits on what the study can tell a general user. The meals came from a controlled ketogenic and standard diet trial, not from restaurant plates, home cooking across varied cuisines, or the mixed dishes most people photograph. The sample of 102 meals is small for a technology used by millions. And app algorithms change, sometimes frequently, so an accuracy measurement is a snapshot rather than a permanent property of the product.
What the study does support is a pattern that is consistent across four independent products rather than a flaw in one. The apps estimated carbohydrates more consistently than fat or total calories. MyFitnessPal and Lose It! estimated higher calorie meals more accurately than lower calorie ones. The largest errors clustered in high-fat, low-carbohydrate dishes.
That last detail is the most usable finding in the paper. Fat carries more than twice the calories per gram of protein or carbohydrate, and it is frequently invisible in a photograph. Oil absorbed into a stir fry, butter in a sauce, dressing already tossed through a salad, and the fat content of a particular cut of meat are all things a camera cannot see. A dish that looks identical can differ by hundreds of calories depending on how it was cooked.
Who Should Pay Attention and Who Should Not Worry
The people most affected are those using photo logging as their primary or only method of tracking, and specifically those who never adjust the app's portion estimate or enter amounts manually.
That group includes people tracking for weight management who cannot understand why the numbers are not moving, people on higher-fat eating patterns including ketogenic diets, and people who eat a large share of prepared or restaurant food where cooking fat is unknown. It also includes anyone using an app's output to make a clinical decision, such as people with diabetes estimating carbohydrate for insulin dosing, though carbohydrate was the most consistently estimated nutrient in this analysis.
People who use these apps for general awareness rather than precision are not really affected. Neither are people who weigh or measure their food and enter it manually, which remains the more accurate approach. The researchers suggested that combining photo-based tools with traditional dietary assessment methods could improve accuracy in everyday use.
What Readers Can Do Now
The practical response is to treat a photo estimate as a starting point rather than a final number.
After photographing a meal, check whether the app has correctly identified every component and adjust the portion size, which is where much of the error originates. Enter cooking fats explicitly, since oil and butter are the single most commonly missed items and the ones carrying the most calories. For meals you eat regularly, weigh the ingredients once and save the entry, which removes the guesswork permanently for that dish. For packaged food, use the barcode scanner rather than the camera.
Anyone using an app output for a medical decision, particularly insulin dosing, should discuss the app's limitations with their clinician rather than adjusting anything on their own.
One further note. Calorie tracking is a useful tool for many people and a harmful one for some. If tracking has started to feel compulsive, or if finding out the numbers may have been wrong produces significant distress, that is worth raising with a clinician. The National Alliance for Eating Disorders operates a helpline staffed by licensed clinicians.
What Happens Next
The next meaningful step is publication. Until the analysis appears in a peer-reviewed journal, the estimates of 250 to 345 calories and roughly 30 grams of fat should be understood as figures presented at a conference rather than settled results.
The app developers have not published responses to this specific analysis. Independent evaluation of consumer nutrition technology remains limited, and no federal accuracy standard photo-based calorie estimation must meet, because these products are marketed as general wellness tools rather than as medical devices. MedicalDaily will report the peer-reviewed publication and any response from the companies named.
The confirmed finding is that four widely used photo-based apps underestimated calories and fat by about a third against meals of known composition. The people most affected are those relying on photo logging alone without adjusting portions. The most reasonable action is to correct portions and enter cooking fats manually. The central uncertainty is whether these figures hold up in peer review and across a broader range of everyday meals.
Frequently Asked Questions
Which apps were tested? MyFitnessPal, Lose It!, Cal AI, and Appediet. All four underestimated calories and fat.
How large was the error? Calories were underestimated by roughly 250 to 345 per meal on average, and fat by about 30 grams. That works out to about one third of the actual content.
Has this research been peer reviewed? No. It was presented as a conference abstract at NUTRITION 2026. Abstracts were reviewed and selected by an expert committee but have generally not completed the full peer review required for journal publication. The results are preliminary.
Why do the apps miss fat specifically? Fat carries more calories per gram than protein or carbohydrate and is frequently invisible in a photograph. Cooking oil, butter, dressing, and the fat content of a cut of meat cannot be seen by a camera.
Does this mean the apps are useless? No. The researchers suggested pairing photo tools with traditional tracking methods. Carbohydrate estimates were the most consistent of the nutrients measured.
How can I make my tracking more accurate? Correct the app's food identification, adjust portion sizes, enter cooking fats explicitly, scan barcodes for packaged food, and weigh and save entries for meals you eat often.
Should I change my insulin dose based on an app estimate? Do not change any medication based on an app. Discuss the limitations of photo based estimation with your clinician.
