How Accurate Are AI Calorie Trackers? What the Research Actually Says

How accurate are AI calorie trackers? Short answer: useful, but less accurate than the marketing suggests. The studies we could find show photo-based apps are good at naming foods and much worse at judging how much is on the plate, especially when fat is hidden in cooking or sauce. In a preliminary 2026 study by NIH researchers, presented as a conference abstract and not yet peer reviewed, four popular apps all underestimated meal calories, by about a third on average.

A quick disclosure before the details: I’m Milos Radojcic, founder of Lettuceburn, an AI photo calorie tracker. That’s a good reason to read this with some skepticism, so every claim below links to its source. Where we talk about our own app, we’ll stick to how it works, not how accurate it is. We haven’t measured that, so we won’t put a number on it.

How accurate are AI calorie trackers? The short answer

Here’s what the research supports, as of October 2026:

• Food identification is fairly good. In a lab test of seven commercial apps, the best identified 90–97% of food components in photos. The worst identified under half.
• Calorie estimates are much less reliable. The same apps that named foods well still missed energy by wide margins, in both directions, especially on mixed dishes.
• The error usually runs low. The newest test found underestimation across the board, with fat you can’t see, like butter, oil and sauces, as the biggest gap.
• Correcting the estimate seems to help. In one company-run study, correcting the AI’s guess brought energy in line with the reference.

None of this means AI photo logging is useless. It means a photo result is a first draft, not a lab measurement.

What the studies actually tested

Research on AI food tracking is young, and the studies don’t all measure the same thing. Here are the ones worth knowing, with what each one actually tested.

Study	What was tested	Main finding
Li et al., 2024, Nutrients	7 commercial apps with AI photo recognition, 22 photos, controlled lighting	Good food identification for the top apps, poor and unpredictable energy estimates for mixed dishes
NIH/NIDDK conference abstract, NUTRITION 2026 (preliminary, not yet peer reviewed)	MyFitnessPal, Lose It!, Cal AI, Appediet on 102 meals with weighed ingredients	All four underestimated calories, by about a third
Fridolfsson et al., 2025, Current Developments in Nutrition	ChatGPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro on 52 food photos	Mean energy error around 36% for the two better models; underestimation grew with portion size
Sasaki et al., 2022, JMIR Formative Research	One Japanese image-recognition app, 120 meals; funded by the app’s maker	Automatic estimates ran low; after people corrected them, energy matched the reference

Commercial apps in a controlled lab

The most detailed look at commercial apps comes from a University of Sydney team. Li and colleagues photographed 22 foods and drinks under ideal conditions: good lighting, a plain background, a 45-degree angle, 30 cm from the plate. They uploaded the same photos to seven apps.

For naming foods, MyFitnessPal identified 97% of the 39 food components, Fastic 92% and HealthifyMe 90%. Lose It! and FatSecret each identified 46%.

Calorie estimates were another story. Only four apps estimated energy automatically, and on single-component foods their average error ranged from −47% (Foodvisor) to +44% (Fastic). Mixed dishes were harder and less predictable. MyFitnessPal underestimated "eggs on toast with butter" by 35% because it missed the butter. Pearl milk tea was underestimated by up to 76%. Yet some Asian dishes went the other way: MyFitnessPal overestimated bibimbap by 270%, and all four apps overestimated beef pho. Errors don’t only run low.

Two caveats matter. The researchers used the free versions available in Australia in 2024, and the photos were taken under lab conditions. The authors note that everyday photos, with messier lighting and angles, could produce larger errors.

Real meals from an NIH metabolic kitchen

The most recent study is also the most direct, but it’s preliminary. A team at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) ran photos of 102 meals through four apps: MyFitnessPal, Lose It!, Cal AI and Appediet. The meals came from a controlled feeding study, so every ingredient was weighed, and the researchers calculated each meal’s nutrition from national food-composition tables. The meals averaged 918 calories.

According to the abstract presented at NUTRITION 2026, every app underestimated calories, by about 33% overall. All four missed about 30 grams of fat per meal. Carbohydrate estimates were the most consistent; fat and protein were not.

This was a conference abstract. The press release says the findings should be treated as preliminary until a peer-reviewed paper is published. It’s still the closest thing we have to an independent test of today’s best-known photo trackers. We look at the per-app numbers for one of them in is Cal AI accurate?

General-purpose AI models

Some newer apps, including ours, use general-purpose AI models rather than a custom-trained food recognizer. A Swedish team tested three of them. Fridolfsson and colleagues analyzed 52 photos of weighed meals and components, some shot at three portion sizes, and asked ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro to estimate them.

ChatGPT-4o and Claude 3.5 Sonnet both had a mean absolute error of 35.8% for energy. Gemini 1.5 Pro’s errors were much larger, at 64.2%. Every model underestimated more as portions got bigger. Part of that came from the setup: vegetables sat at the front of the plate, so on larger portions the rice, pasta and protein behind them were harder to see.

The authors concluded the two better models were about as accurate as traditional self-reported food diaries, without the effort, but "not yet suitable for precise dietary assessment" where exact numbers matter. These tests used 2024 model versions and one simple prompt, so they don’t tell you how any specific app performs today.

What happens when people correct the AI

One older study looked at the step most people skip: fixing the result. Sasaki and colleagues tested a Japanese app called CALO mama on 120 prepared meals. The automatic estimates underestimated energy and matched the reference for 11 of 30 nutrients. After research staff corrected food names and portions by eye, energy matched the reference, and 29 of 30 nutrients did.

Worth knowing: the study was funded by the company that makes the app, and two of the four authors, including the first and corresponding author, work there. The correctors were research staff, and the study didn’t measure how long corrections took. So treat it as one encouraging result, not proof. It does fit the pattern in the other studies, where most of the error sits in things a person can spot, like a missing pat of butter.

Why AI calorie tracker estimates miss

The studies point to the same few causes again and again.

Hidden fat, sauces and drinks

A camera can’t see the tablespoon of oil in the pan or the butter melted into toast. According to USDA FoodData Central, a tablespoon of olive oil (13.5 g) carries about 119 calories. That fits the NIH team’s finding that fat was the biggest gap, and Li et al. traced MyFitnessPal’s 35% miss on eggs on toast to the butter. Drinks with hidden sugar, like milk tea, cause a similar problem.

Why photo calorie counting struggles with portions

A photo is two-dimensional. The AI has to guess depth, density and weight from things like plate size and utensils. A deep bowl of pasta and a shallow one can look nearly the same from above. In the Fridolfsson study, models struggled to scale up their estimates as portions grew, so bigger meals were underestimated more.

Mixed and less familiar dishes

Stews, curries, noodle soups and layered dishes hide their ingredients. In the Li study, apps often failed to pick out the parts of Asian mixed dishes, and their estimates swung from far too low to far too high. Training data matters too. A model that has seen far more burgers than bibimbap will guess better on burgers.

The reference numbers themselves

Even a perfect guess about grams has to be converted into calories using a food database or food-composition knowledge, and those values are averages. Two chicken thighs from different farms aren’t identical. Even the NIH study’s reference values came from food-composition tables, not lab measurements of each plate.

Accuracy isn’t only an AI problem

A fairer comparison than perfection is the other ways people log food. Manual logging depends on picking the right database entry and guessing portions, and the Li study found that, on average, 16 manual-logging apps overestimated a Western diet by about 1,040 kJ (about 250 kcal) a day and underestimated an Asian diet by about 1,520 kJ (about 360 kcal) a day. FDA rules give nutrition labels some slack, barcode scans depend on the database entry behind them, and self-reported food diaries tend to run low. We cover the manual side in how accurate calorie counting apps are, and compare the two logging methods in barcode scanner calorie accuracy vs photo logging.

The NIH researchers made the same comparison: in their preliminary results, AI photo apps underestimated by about as much as self-reported food diaries typically do. So AI photo logging isn’t uniquely bad. It’s another estimate with its own blind spots, and it’s faster.

How to read an app’s accuracy claim

App homepages make claims like "about 80% accurate" (Cal AI) or "2x more accurate than nutritionists" (SnapCalorie). Before you take one at face value, ask three questions:

1. Accurate at what? Naming foods and estimating calories are very different tasks. An app can name 97% of foods and still miss calories badly.
2. Measured how, and by whom? A company’s own test on its own dataset is not the same as an independent study with weighed meals.
3. On which meals? Single foods under studio lighting flatter every app. Mixed dishes, restaurant plates and dim kitchens are where errors grow.

We apply these questions to the wider field in our comparison of the best AI calorie trackers. For trackers in general, including ones without AI, see our broader calorie counter app comparison.

How to get better numbers from any photo tracker

You can’t make a camera see oil, but three habits close much of the gap:

• Tell the app what it can’t see. "Cooked in 1 tbsp olive oil." "Full-fat Greek yogurt." "Two slices, not one."
• Check the ingredient list, not just the total. If the butter is missing, add it.
• Use the label or barcode for packaged food instead of a photo of the food itself.

There’s a full ten-habit checklist, covering angles, lighting and size references, in how to log food with photos accurately.

How Lettuceburn handles error

We built Lettuceburn assuming the AI will sometimes be wrong, because the research says it will be. Here’s what that looks like in practice, without an accuracy number attached.

• Up to three photos per meal. Shoot the plate from above and the side. The AI is told the photos show the same meal, to count each food once, and to use the extra angles to refine portions and spot hidden food.
• An optional text hint. A short note like "that’s oat milk" or "fried in butter" is passed to the AI as information about the meal.
• An allowance for fat it can’t see. The AI is told to allow for cooking oil when a dish looks sautéed or fried, and to lean higher on fat for restaurant plates. That doesn’t solve the hidden-fat problem, but it pushes against the direction the studies found.
• Totals come from the ingredients. The AI breaks the meal into up to eight ingredients, each with an estimated weight in grams. Lettuceburn adds up the meal total from that list. The AI never reports a total on its own, so the list you check and the number you log can’t disagree.
• Fix Results before you save. Tell the AI what’s wrong in one sentence ("that’s brown rice, and there are two chicken thighs") and it re-analyzes the meal, treating your correction as fact. You can also edit any ingredient’s grams directly, add a missing item, or delete one.
• An honest "not sure." When the AI reports low confidence, the result says so and suggests checking portions. We never show confidence as a badge of accuracy.
• Better tools for packaged food. The nutrition-label scan on our features page reads the per-serving column printed on the panel instead of guessing, and the barcode scanner looks products up in Open Food Facts.

One more honest note: on the free lane you bring your own AI key. With an OpenRouter key you can also pick the model, so results partly depend on that choice. You can see the full flow in how Lettuceburn works, and the two ways to pay for AI (or not) on our pricing page: Pro, or bring your own key, free forever.

When precision matters more

For most everyday goals, a consistent log beats a perfect one. If you log the same way every day, your weight trend over a few weeks tells you whether your estimates are running high or low, and you can adjust. Our guide to how many calories you should eat explains how to set and adjust a target from your own numbers.

Some situations need more precision than any photo app can promise. If you count carbohydrates to dose insulin, manage a kidney condition, or work with a clinician on a medical diet, use the methods your care team recommends. The Li study specifically flagged carbohydrate errors as a risk for people with diabetes.

The bottom line

AI calorie trackers are good at recognizing food and much less good at judging how much fat and food is actually on the plate. The studies we found mostly show them running low, sometimes by hundreds of calories per meal, and occasionally far too high on mixed dishes. Used as a fast first draft that you check and correct, they’re a reasonable way to keep a food log. Used as a precise measurement, they’ll mislead you.

So pick an app that shows its work, lets you fix it before saving, and doesn’t hide the estimate behind a confident number. Then spend ten seconds checking each meal. Of everything in this guide, that check is the habit the research most consistently points to.