Why does my calorie tracking app underestimate calories?

Because it is guessing. A photo or video app has to identify what is on your plate, judge the portion size, and estimate hidden ingredients like oil, dressing, and sauce, all from an image. Fat and mixed dishes are the hardest to judge, so error tends to run in that direction. Even in lab tests with weighed, known meals, automated identification and energy estimates carried real error.

Is MyFitnessPal accurate for tracking calories?

It depends on how it is used. In a 2024 randomized crossover lab study, MyFitnessPal's manual search and log method estimated energy intake within about 7 percent of weighed food, better than a portion size template app tested the same way. In a separate study of adults with obesity, MyFitnessPal underestimated energy, carbs, and fat versus dietitian coded records. Manual logging can be reasonably close, but it is still an estimate, not a measurement.

Are AI photo calorie trackers accurate?

Not consistently. In the first peer reviewed validity study of an AI photo and video food identification app, automated food identification exactly matched the true food only 46 percent of the time across meals, ranging from 27 percent on a mixed dish up to 67 percent on a simple one. Automated energy estimates carried up to 43 percent error when drinks were included, and error was larger for higher calorie meals.

If the numbers are unreliable, should I stop logging my food?

No. Logging is still one of the most consistent predictors of weight loss success in the research, because the value comes from the awareness and pattern recognition, not from decimal point accuracy. The fix for an inaccurate estimate is not to abandon tracking, it is to treat the number as a starting point you correct, not a verdict you accept.

Back to Blog

Why Your Calorie Tracking App's Estimate Is Probably Off

Photo and app based calorie estimates carry real, measured error, even in controlled research settings. Here is what the peer reviewed studies actually found, and why correcting the number matters more than the number itself.

Heather Bauer, RD · October 22, 2026

Your calorie tracking app is not lying to you. It is guessing, and the peer reviewed research on exactly how much it guesses wrong is more specific than most people realize: automated food identification from a photo or short video matches the true food only about half the time, and automated energy estimates can be off by double digit percentages, worse for mixed dishes and higher calorie meals.

That is not a reason to throw out the app. It is the reason a number on a screen should be a starting point you correct, not a verdict you accept.

What the research actually found

How far off is AI photo food identification?

The first published validity study of an AI based food identification app (Lozano et al., Current Developments in Nutrition, 2023) had 24 adults eat known, weighed lab meals and then log them using an app that identifies food automatically from a short video, no barcode or reference marker needed. The automated system found an exact match to the true food only 46 percent of the time. Accuracy swung hard by meal: 27 percent exact matches on a mixed pizza meal, up to 67 percent on a simple hamburger meal. Automated energy estimates carried up to 43 percent error when beverages were included, dropping to 16 percent once drinks were excluded, and error was consistently larger for higher calorie meals.

Translation: the easier the plate is to “read,” the closer the number gets. A burger and fries is a more solvable image problem than a mixed casserole, a smoothie, or a restaurant dish with hidden oil and sauce.

Does manual logging do any better?

Sometimes, and it depends on the app and the comparison. A 2024 randomized crossover lab study (Apolzan et al., American Journal of Clinical Nutrition) had 43 adults eat weighed lab meals, then log them two ways: a portion size template app, and MyFitnessPal’s manual food search. MyFitnessPal’s energy estimate came out 7.0 percent over the true, weighed value, close enough to count as statistically equivalent to weighed intake. The portion size app underestimated energy by 13.3 percent and did not reach that bar. A separate study of MyFitnessPal used by adults with obesity in the Philippines, checked against dietitian coded food records, found the app tended to underestimate energy, carbohydrate, and fat, and overestimate protein.

Put those two findings next to each other and the honest read is: manual logging in an app like MyFitnessPal can land close to the true number, but “close” is doing real work in that sentence, and which direction the error runs depends on the study, the population, and the food.

Why is fat the hardest thing to estimate?

Across these studies, fat is consistently the macronutrient apps struggle with most, whether the app is guessing from a photo or a person is estimating portions by eye. Fat calories are dense and often invisible: the oil a restaurant cooks in, the dressing mixed through a salad, the butter on a pan. A gram of fat carries more than double the calories of a gram of carbohydrate or protein, so a small miss on fat becomes a large miss on the total.

Why this is not an argument against tracking

None of this means logging your food is pointless. It means the opposite. Decades of self-monitoring research, including trials like the Kaiser Permanente Weight Loss Maintenance study, tie consistent food logging to significantly better weight outcomes, and that benefit comes from the awareness of writing it down, not from the calorie count being perfect to the gram. Nothing here changes that.

What it does mean is that a single AI generated or app generated number deserves a second look, not blind trust. If a plate of pasta with a cream sauce logs at 400 calories, that is a plausible place for the estimate to be running low, not a reason to panic, and not a reason to stop logging either. That’s just data. Notice it, correct it if you can, and keep going.

The fix is correction, not perfection

This is the actual argument for a coaching layer sitting on top of a calorie estimate instead of the number standing alone. An estimate that gets checked, questioned, and adjusted by a person, or a system built to catch the patterns above, gets more useful every time it is corrected. An estimate nobody ever double checks just keeps being wrong in the same direction.

Practically, that looks like:

Treating restaurant meals, mixed dishes, and anything with a visible sauce or dressing as more likely to be underestimated, and rounding your expectation up rather than down.
Logging the parts of a meal you are unsure about separately (dressing on the side, oil used in cooking) instead of letting the app guess at them inside one photo.
Using the trend over weeks, not any single day’s total, to judge whether your plan is working. One off estimate does not change the pattern.
Correcting an estimate when you know it is wrong. A quick fix today is also the fastest way to make tomorrow’s estimate better, whether a human or a system is doing the adjusting.

If you want that correction built into your day instead of left to chance, that’s exactly what a coaching layer around your logs is for: not a more precise camera, but a second pass that catches what the first guess missed.

The bottom line

Your app’s calorie number is an estimate, and the peer reviewed research says it is a specific kind of estimate: reasonably close for simple, easy to photograph meals logged carefully, and meaningfully further off for mixed dishes, restaurant food, and anything hiding fat. Keep logging. Just treat the number as a first draft, not a final answer, and correct it when you know better.