Research

How accurate are they, really?

Photo-based calorie estimates are wrong in a predictable direction: low. This page collects what the measured evidence says about how far off they run, why portion size and fat are the hardest parts, and whether a daily shortfall averages out. Every figure is linked in place.

By Alen Yaco Double bachelor's in biological sciences and social ecology, UC Irvine Published 18 September 2026

Not on this page. How CalorieCounter labels its own uncertainty, using a confidence word rather than a percentage, is covered at /ai-calorie-counter/ and argued in full at /blog/three-words-instead-of-a-percentage/, not on this page. The app is unreleased and cannot be installed yet.

How accurate are AI calorie apps?

They run low rather than randomly wrong. A 2020 meta-analysis of 13 studies put the shortfall at 179 calories a day, and 448 against doubly labelled water. Preliminary 2026 research on 4 photo apps found weighed meals underestimated by 250 to 345 calories each, about a third.

No single figure describes the category. Any app publishing one is averaging its good days with its bad ones.

The peer-reviewed anchor is a 2020 meta-analysis in Clinical Nutrition: 13 studies, 606 participants, with intake under-reported by 179 calories a day and by 448 against doubly labelled water. That method measures expenditure from isotope turnover rather than asking what anyone ate, which is why it finds the larger gap.

The cleanest recent test comes from the National Institute of Diabetes and Digestive and Kidney Diseases. Researchers led by Aaron Hengist photographed 102 meals prepared in a metabolic kitchen, where every ingredient was weighed to 0.1 gram, then ran the standardized photos through MyFitnessPal, Lose It, CalAI and Appediet. All 4 apps came in low, by 250 to 345 calories and about 30 grams of fat per meal, according to the American Society for Nutrition. The release calls that about a third and prints no precise percentage. Neither should anyone quoting it.

Little else in this field compares as cleanly. The caveat is the format. It was presented at NUTRITION 2026 in July, and abstracts there were selected by a committee of experts but did not go through the peer review a journal requires, which the American Society for Nutrition says on its own release. No paper exists yet. Treat the figures as preliminary: 4 apps, tested once, on one kitchen's meals.

Evidence	Compared against	Reported gap
Ho et al., Clinical Nutrition 2020 (13 studies, 606 people)	All methods pooled	-179.32 kcal (95% CI -269.50 to -89.15)
Ho et al., same review	Doubly labelled water	-448.04 kcal (95% CI -755.52 to -140.56)
NUTRITION 2026 (102 meals, preliminary)	Weighed ingredients	250 to 345 kcal and about 30 g fat per meal

One figure is per meal against a kitchen scale, the other per day against a metabolic reference. They agree on direction and disagree on size. Read the intervals too: heterogeneity was 89% for the pooled estimate and 95% for the doubly labelled water comparison, so the studies disagreed sharply. One result also cuts against the simple story. The review found no significant difference for protein or fat, only for total energy.

Why is portion size the hardest part for a calorie app?

Portion size is the weakest link. When 7 commercial food-recognition platforms were tested on 185 real food photos in 2020, none could estimate portion size at all. A review in the International Journal of Obesity that year called automatic portion estimation immature, not solved.

Identifying a food and quantifying it are different problems, and the second is harder. The head-to-head test in JMIR Formative Research ran 185 photos through 7 commercial platforms and reported that not one returned a portion size. They could guess the food. They could not say how much was there.

Volume is the difficulty. A photograph flattens depth, so height has to be inferred from one flat view. Across the 52 studies in the Annals of Medicine review, average relative errors for volume ran from 0.09% to 33%. The best figure anywhere, 9.4% for the research system GoCarb, is restated by a 2025 scoping review in Frontiers in Nutrition. GoCarb put a reference object beside the plate, the crutch a bare photo lacks.

That platform test predates current multimodal models. Still, automatic portion estimation was called immature by the International Journal of Obesity, and nothing published since has shown it solved.

Which foods do AI calorie apps get most wrong?

Mixed dishes. A 2023 systematic review in Annals of Medicine covering 52 studies found relative errors were smaller for photos of one simple food than for photos holding several. Curries, stews, salads and loaded plates hand the model overlapping items and unclear boundaries.

The mechanism is segmentation. Before an app estimates anything it has to decide where one food ends and the next begins. A single apple offers few distractions. A curry offers overlapping foods at different heights with no clean edges, and every wrong boundary feeds the calorie number.

The Annals of Medicine review screened 14,059 publications and included 52 from 2010 to 2023. Average relative errors, and their spread, were smaller for single-food images than for multiple-food images.

The size of that category is the part usually left out. In the 185-photo platform test, 72 images were mixed dishes, the largest group, ahead of 47 beverages, 36 processed foods and 30 plain single foods. The hard case is not an edge case. What follows from that when you are holding the phone is set out separately.

Why do AI calorie apps miss fat?

Fat is frequently invisible in a photo. Oil absorbed during cooking, butter stirred into a mash and dressing already tossed through a salad change the calorie count without changing the picture. The NUTRITION 2026 work found the apps missed about 30 grams of fat per meal.

The 2020 review in the International Journal of Obesity names this failure mode. Ambiguous items such as diet against regular soda, cooking methods such as fried against baked, and hidden ingredients such as butter in a mash usually need the person eating to annotate the photo. A camera cannot recover what a camera cannot see.

The NUTRITION 2026 work supplies the magnitude and a pattern. In a secondary analysis covering more than 200 further meals, the apps did worst on low-carbohydrate ketogenic meals, which the researchers attribute to higher fat being consistently underestimated. Carbohydrates were the most consistently estimated of the 3 macronutrients.

These apps tend to underestimate calories, especially from fats.

That is Aaron Hengist, in the American Society for Nutrition release.

The record has a gap of its own. The 2025 Frontiers in Nutrition scoping review compiled accuracy figures for 10 named systems, and none reported a metric for hidden fats or oils. The largest named blind spot in the field rests on one preliminary abstract. The everyday version is covered separately.

Is typing your food in more accurate than photographing it?

Not reliably. The Clinical Nutrition meta-analysis found no statistical difference between image-based assessment and traditional methods such as diaries and recalls, except against doubly labelled water. Its authors concluded that image-based methods carry serious measurement error in the same way traditional methods do.

This is where most writing about app accuracy goes wrong. The useful comparison is not photo against perfect. It is photo against whatever you would otherwise do.

A 2017 study in Nutrients dosed 45 adults aged 21 to 65 with doubly labelled water and had them keep an image-based food record for 7.5 days. Reported intake and measured expenditure correlated at 0.58, their mean ratio was 0.81, and 53% of participants were under-reporters by a mean of 563 calories a day. The authors judged that record comparable to traditional dietary records.

The shortfall is not a property of cameras. It is a property of people measuring their own eating, and it turns up in diaries, recalls and photos alike.

Why do published accuracy numbers vary so much?

Because they measure different things. Research benchmarks report food classification accuracy from 49.19% to 99.05% depending on the dataset. Commercial platforms tested on ordinary food photos scored far lower, from 9.1% to 62.9% for a correct first guess. Neither figure describes the app on your phone.

Benchmark accuracy on curated research datasets is the flattering kind. A 2022 review in Advances in Nutrition included 78 studies, 45 of them using deep learning, with classification accuracy from 49.19% to 99.05% across datasets.

Real photographs through real commercial platforms look very different.

Platform	Top-1 accuracy	Top-5 accuracy
Calorie Mama API	62.9%	87.6%
Bitesnap	48.9%	71.0%
Foodvisor	46.2%	71.5%
Clarifai	38.2%	64.0%
IBM Watson	25.3%	43.5%
LogMeal	24.2%	44.1%
Google Vision API	9.1%	24.2%

Those come from the 185-image JMIR Formative Research test, which predates current multimodal models. Read it as the distance between a benchmark score and a real photograph.

Calorie error spans a similar range. The Annals of Medicine review found average relative errors from 0.10% to 38.3%, and single errors from 0.00% to 79.6%. It could not pool them, because databases and reporting varied too much, and concluded the tools need more development.

Does a daily underestimate cancel out over a week?

No. Random error averages out over days, and a bias does not. The pooled estimate in Clinical Nutrition is negative across its whole confidence interval, and all 4 apps in the NUTRITION 2026 test missed in the same direction. That is a lean, not noise.

The difference decides what a week of logs is worth. Noise is an estimate 200 calories high on Monday and 200 low on Tuesday. Average 7 of those and the error mostly disappears. Bias is an estimate 200 low every day. Average 7 of those and you are still 200 low, and 1,400 calories short across the week.

Everything above points at bias. The pooled 179 calories a day is about 1,250 a week. The doubly labelled water figure of 448 a day is about 3,100 a week. Averaging more days does not shrink either one.

So more logging buys precision rather than accuracy. It tells you what your lean is, without removing it. The figure worth reading is the change in your own totals, not any single day, because a steady bias cancels when you compare one week against the last.

That holds only while the bias stays put, and it moves when your eating moves. The shortfall was worst on the highest-fat meals, so 2 different weeks of eating are not measured with the same ruler.

What should you do with the number an app gives you?

Treat it as an estimate with a known lean. The measured evidence says photo estimates come in low, and worst for fat, mixed dishes and portion size. Scanning a barcode where one exists, and reading the number as a floor rather than a midpoint, both follow from that.

Two things follow from the research above, and neither is changing app:

Scan the barcode when one exists. A packaged food carries its own numbers, which takes the photo out of the step that fails.
Read the output as a floor, not a midpoint. Every study here found the error sitting under the true value, not scattered either side of it.

Weighing the few dense staples you eat most often, once, closes more of the gap than switching apps does, and that calibration is a separate subject.

Two of the remaining fixes have pages of their own: the calories a photo cannot show, and how to shoot a plate so there is less to guess.

The last thing is to distrust a headline accuracy percentage. One number averages the meals a tool handles well with the meals it handles badly, and says nothing about the plate in front of you. That is the case for a confidence word in place of a percentage, the approach CalorieCounter takes, though it is unreleased and cannot be installed yet. What it does is here, and the reasoning is here.

What the record supports is narrow: a direction, from 4 apps in 1 kitchen and 13 studies of image-based records, and a size that more than doubled depending on the reference used. No published figure is precise enough to describe your own plate.

FAQ

Is a 90% accuracy claim from a calorie app believable?

Ask what was measured. Naming a food and counting its calories are different tests, and a top-5 score counts a guess as correct if the right answer appears anywhere in 5 tries. Published figures run from single digits to near-perfect.

Does a better phone camera make calorie estimates more accurate?

Not by much, on the published evidence. The named failure modes are not resolution. They are separating overlapping foods, judging depth from a flat image, and accounting for ingredients that never appear in it. A sharper photo of a curry still will not show how much cream went in.

Do newer AI models fix this?

Not on the evidence so far. Most head-to-head testing of commercial platforms dates from 2020 and predates current multimodal models. The one recent test of consumer apps, presented in 2026 and unpublished, found the same direction and a shortfall of roughly a third per meal.

Sources

Be first in line.

CalorieCounter is coming to the App Store for iPhone. Leave your email and you get the link the day it is live. One email, nothing else.