Short answer: they are useful for getting a rough number fast, and they tend to undercount. In July 2026 a team at the National Institutes of Health tested four popular photo calorie apps against meals whose every ingredient had been weighed. On average the apps missed about a third of the calories and a third of the fat.
The results are early. They were presented as a conference abstract and have not yet been through full peer review, so treat the exact numbers as provisional. The direction of the error is the part worth acting on now.
What the researchers did
The work came from the intramural research program of the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), part of the NIH. Olivia Charles presented it on 25 July 2026 at NUTRITION 2026, the American Society for Nutrition’s annual meeting.
The meals came from a separate NIH study on ketogenic versus standard diets, which meant they were cooked in a metabolic kitchen where ingredients are weighed to the nearest 0.1 gram. That gives a reference most app tests never have: the true calories of the plate.
The team photographed 102 of those meals in a standard way and gave the photos to MyFitnessPal, LoseIt!, Cal AI and Appediet. They later ran more than 200 additional meals to look at what made the estimates better or worse.
What they found
| Result | |
|---|---|
| Calories | 250 to 345 kcal too low per meal, on average across the four apps |
| Overall | Calories and fat both about one third too low |
| Fat | About 30 g too low per meal |
| Carbohydrates | The most consistent of the three macronutrients |
| Keto meals | Early data suggests the apps did worse on these high-fat meals |
MyFitnessPal and LoseIt! were more accurate on larger, higher-calorie meals than on smaller ones. The press release does not give a per-app ranking, so nobody can honestly tell you from this study which of the four “won”.
To see why the miss matters, take a plan built on a 500 kcal daily deficit. If a gap of around 300 kcal showed up at each of three logged meals, the deficit you think you have would be gone, and the scale would stall while the app says you are on track.
Why a photo misses fat
Fat is the most calorie-dense thing on a plate, at 9 kcal per gram, and it is the hardest to see. The 30 g the apps missed per meal is a little more than two tablespoons of oil, roughly 270 kcal. Oil soaks into rice and vegetables, butter melts into a sauce, and dressing coats the leaves. In the photo they look like nothing.
A photo also flattens depth. The model has to guess how deep the bowl is and how densely the food is packed, and two plates that look identical from above can differ by a few hundred grams.
Carbohydrate estimates held up better in the study, which fits: bread, rice and pasta are fairly visible and their calories per gram vary less.
How to log around it
The researchers’ own advice was to use photo apps alongside traditional methods, and to be careful if you never adjust portions or enter amounts yourself. In practice that means a few habits:
- Tell the app about the fat you cannot see: “fried in oil”, “with butter”, “creamy sauce”, “dressing on top”. One tablespoon of oil is about 125 kcal.
- Correct the portion when you know it. If the package says 150 g, or you ate half, say so.
- Weigh a few of your usual meals once. You learn what your normal bowl of pasta actually weighs, and every later photo of it gets easier to judge.
- Judge the plan by your weekly average weight over two or three weeks. If the trend disagrees with the log, believe the trend and adjust your target.
Where ChatMyFood stands
ChatMyFood was not part of this study, so we cannot claim it would have scored better. What we can say is how it is built in response to this kind of error. Every photo estimate comes with a range, such as 620 ±80 kcal, so you can see when a number is shaky. When the oil or the portion would change the answer a lot, it asks you a short question instead of guessing. And you can correct anything by typing, for example “only half” or “no dressing”.
None of that makes a photo as accurate as a kitchen scale. It makes the uncertainty visible, which is what this study suggests users need.