AI calorie counter accuracy: how reliable is counting calories from a photo?

Snap your plate instead of searching for every food in a database: that is the promise of MyFitnessPal (Meal Scan), Lose It! (Snap It), Cal AI, Foodvisor, SnapCalorie and Ember, the app that publishes this guide. Behind the button, the process is almost always the same, in three steps. The model first recognizes the foods, then estimates how much of each is on the plate, and finally converts those grams into calories and macros using a food composition database.

The first step has become very reliable. In an Irish study published in 2025, ChatGPT identified the foods in 114 meal photos with 93% precision. The second step is far shakier: a photo is a flat image with no scale and no tape measure. The model has to guess the thickness of a steak, the depth of a bowl or how much oil a pan of stir-fried vegetables soaked up. The third step depends on what was recognized: if the app sees plain white rice where there is fried rice, the calculation starts off on the wrong foot.

So a high recognition rate tells you nothing about calorie accuracy. An app can name every item on the plate perfectly and still be badly off on the portion. When an app listing quotes an “accuracy” percentage, ask yourself whether it measures how well dishes are identified or how close the calories actually are. To compare features across the main apps, see our roundup of the best calorie counting apps.

Independent studies are still rare, but they have multiplied since consumer vision models arrived. The protocols look alike: meals weighed to the gram are photographed, run through an app or a model such as ChatGPT, and the estimate is compared with the true value.

The most talked-about result comes from researchers at the NIDDK, the NIH institute for diabetes and nutrition. Presented in July 2026 at NUTRITION, the American Society for Nutrition’s annual meeting, their study ran 102 meals prepared in a metabolic kitchen through four apps: MyFitnessPal, Lose It!, Cal AI and Appediet. All four underestimated calories by about a third, or 250 to 345 kcal per meal on average, and fat by about 30 g. It is a conference abstract, not yet a peer-reviewed paper, and results were not broken down app by app.

Two mechanisms show up again and again. The first is portion size. In the Irish study, ChatGPT judged small plates correctly but underestimated weight in 76% of photos: a large portion averaging 798 g was estimated at 530 g, so a third of the food vanished from the calculation. The Swedish team that tested ChatGPT-4o, Claude and Gemini saw the same pattern: the bigger the portion, the further the estimate fell short. Models seem to pull every plate toward a “standard” serving, which flatters small eaters and penalizes generous plates.

The second mechanism is fat, the most energy-dense of the macronutrients at 9 kcal per gram. A spoonful of oil soaked into vegetables, a knob of butter melted into pasta or a sauce stirred through a dish is simply invisible. The 30 g of missing fat in the NIH study account for about 270 kcal on their own, most of the measured shortfall. The researchers also noted that the apps struggled more with ketogenic meals, which are very high in fat.

Fat errors do not always point the same way. On fairly lean Japanese hospital trays, AI models actually overestimated fat by more than 20%. The authors call it the “invisible nutrient” problem: since the model cannot see the fat, it applies an average. That works for an ordinary dish, much less for a very rich or very light one.

Some meals are made for photos: a plate with separate, clearly visible items (a fish fillet, rice, green beans) usually gets a good estimate. Trouble starts as soon as the information leaves the frame. In the ChatGPT-5 study, dishes served in bowls produced the largest errors, because the camera cannot see depth. A curry, a soup, a loaded salad or pasta in sauce hides its ingredients in layers.

The camera angle matters too. Shot at a slant, a plate looks smaller at the back than at the front, and perspective distorts the volume estimate. Dim light or a blurry photo get in the way of recognition itself. Above all, an image says nothing about what happened in the kitchen: an oiled pan, butter in the mash, sugar in the tomato sauce, cheese melted under the surface.

Drinks are the extreme case. A glass of dark liquid could be flavored sparkling water, cola or iced tea; an orange one could be juice or soda. For an AI, the difference is often impossible to tell, yet the gap is more than 100 kcal per glass.

Before writing off photo logging, compare it with what it replaces. When people log meals themselves, they forget a cookie, shrink a portion and round down. The gold standard for measuring this is doubly labeled water: you drink water enriched with stable isotopes, and the rate at which they leave the body reveals your real energy expenditure over one to two weeks. In someone whose weight is stable, intake equals expenditure, so their reported intake can be checked.

The results are clear. In the American OPEN study of 484 adults, men underreported their intake by 12 to 14% in 24-hour recalls, and women by 16 to 20%. With a food frequency questionnaire, the gap rose to 31 to 38%. A study published in the New England Journal of Medicine in 1992 followed a small group of people convinced they could not lose weight on a diet under 1,200 kcal: on average they underestimated their intake by 47% and overestimated their physical activity by 51%.

None of this means people lie. Judging a portion by eye is hard for everyone, and small extras are easy to forget. It is often the real reason behind a seemingly mysterious weight loss plateau. Manual logging with every ingredient weighed remains the most accurate method, but it is so tedious that few people keep it up for long. The Swedish researchers point out that the error of ChatGPT-4o and Claude fell in the same range as food diaries of athletes checked against doubly labeled water, without the logging effort.

Good news: for weight management, consistency matters more than perfection. If your app underestimates your meals by 15% every day, in a fairly steady way, your daily total is wrong, but your trends are still readable. You can see which days you eat more, which meals weigh the most and whether your changes are working.

Your scale then acts as the referee. A calorie deficit proves itself over time: if your average weight has not moved after three or four weeks while the app shows a deficit, your meals probably contain more calories than estimated. Rather than cutting portions further, first hunt down the error sources described above: oil, sauces, drinks and big plates.

Accuracy down to the gram only matters in specific cases: medical follow-up, closely supervised athletic preparation, or a condition that requires controlling one nutrient. In those situations, photo estimates replace neither weighing nor the advice of a doctor or registered dietitian. For a fitness or weight goal, a simple method you stick with every day beats a perfect one you drop after ten days.

At a restaurant, assume the dish contains more fat than it looks: professional kitchens are generous with butter, cream and oil for flavor and texture. Snap the plate, then add what you know: “fries”, “butter sauce”, “dressing on the side”. Our tips for eating healthy at restaurants also help you pick dishes that are easier to estimate.

At home, you have a huge advantage: you know the recipe. Note how much oil went into the pan and split it between servings. For a family dish, a description such as “one sixth of a lasagna made with 400 g of beef and 200 g of cheese” will be far more accurate than a photo of the top of the bake. The ChatGPT-5 study shows it clearly: the error fell from 30.5% with the photo alone to 24.4% when details such as the type and amount of fat were specified, then to 13.9% with the full ingredient list.

For packaged food, a photo is not the best tool: the nutrition label gives you the exact value. Scan the barcode whenever you can and keep photos for mixed plates.

These habits take a few seconds and fix most of the errors the studies describe:

Photo calorie apps are very good at recognizing what is on the plate and much less good at judging how much of it there is. From a photo alone, expect an average error of around 25 to 35%, usually on the low side, and larger for big portions, saucy dishes and fatty meals. That is no worse than human memory, which misses 12 to 20% of what we eat on average. Shoot from above, give a sense of scale, describe the oil and sauces, correct the amounts and weigh your food now and then: that is what turns a rough estimate into a reliable tool. Finally, watch your weight trend over several weeks, because it is the real test of whether your numbers hold up.