How accurate are AI calorie trackers?

The short answer

That doesn’t make photo logging useless. It’s fast, and fast logging is logging you keep doing. It just means the number needs a quick look before you trust it, especially for anything cooked in oil or covered in sauce.

What did the 2026 NIH study find?

In July 2026, researchers from the National Institute of Diabetes and Digestive and Kidney Diseases, part of the NIH, presented a test of four photo calorie apps at NUTRITION 2026, the American Society for Nutrition’s annual meeting.

The setup is what makes it useful. The team took 102 meals from a diet study at the NIH Clinical Center. Those meals were cooked in a metabolic kitchen, where every ingredient is weighed to the nearest 0.1 gram, so the true calories were known. Standard photos of each meal went into four apps: MyFitnessPal, Lose It!, Cal AI and Appediet.

What they found:

• All four apps undercounted. On average they came in about 250 to 345 calories too low per meal, about 33% below the meals’ real energy content.
• Fat was the biggest miss. The apps underestimated fat by about 30 grams per meal. At 9 calories per gram, that’s around 270 calories, close to the whole gap on its own.
• Carbs were easier. According to HealthDay’s report on the four apps, the apps estimated carbohydrates better than fat or protein.
• Lighter meals were harder for some apps. MyFitnessPal and Lose It! were more accurate on higher-calorie meals than on lower-calorie ones, as News-Medical reported.
• High-fat meals were worse. A follow-up with more than 200 additional meals suggested the apps struggle more with low-carb ketogenic meals, which tend to be higher in fat.

Healio reported the average undercount for each app:

The point isn’t which app came first. All four missed in the same direction. Aaron Hengist, one of the NIH researchers, said people who use a photo app “without adjusting the portions or entering the amounts of food should take the results with a grain of salt.”

Two caveats. The results are preliminary: they come from a conference abstract that hasn’t been through full peer review yet. And the meals came from one diet study, so other foods and other apps could do better or worse.

Why do photo calorie counters get it wrong?

A photo shows what’s on the plate. It doesn’t show how the food was made. Most misses come from four places.

Hidden fats

Oil in the pan, butter melted into vegetables, dressing tossed through a salad. These add a lot of calories and are often invisible. A tablespoon of olive oil is about 120 calories, according to USDA data. The 30 grams of fat the apps missed is about what’s in a little over two tablespoons of oil.

Portion size from one flat photo

A single photo has no real depth and no scale. A deep bowl hides how much is underneath, and without something of known size in the shot, a small plate of pasta and a big one can look alike. The app has to guess the weight of each food from its shape on the screen, and a small error in a dense food like rice, pasta or cheese is a lot of calories.

Mixed dishes and sauces

Curries, stews, burritos and casseroles hide their ingredients inside. A sauce could be tomato or cream. The camera sees the top layer and has to assume the rest.

Foods that look alike

Whole milk and skim milk look the same in a latte, and regular and diet soda look the same in a glass. White rice and cauliflower rice, or regular and light mayo, are easy to mix up. Only you know which one it was.

Is manual calorie counting more accurate?

Not automatically. Writing down what you eat has its own errors. In the IDATA study of 1,075 adults aged 50 to 74, researchers compared what people reported with doubly labeled water, a lab method that measures how much energy the body actually uses. Four-day food records undercounted energy by 18 to 21%, and online 24-hour recalls by 15 to 17%.

The NIH researchers noted the same thing: the apps’ one-third undercount was similar to what earlier studies found when people reported their own diets. Even packaged food labels have some room. Under FDA labeling rules (21 CFR 101.9), a packaged food’s calories can run up to 20% above the label before the food counts as mislabeled.

So every method is an estimate. What matters is how big the error is and whether it always leans the same way. A steady miss adds up: 300 calories short at each of three meals is 900 calories a day your log doesn’t show.

How can you get a better estimate from a photo app?

These habits help with any app, including ours.

• Get the whole plate in the frame. Include the edges of the plate or bowl, so there’s a sense of scale.
• Use good light and a slight angle. A shot from a little above the side shows how tall a pile is, which a straight-down photo hides.
• Keep foods apart when you can. Sauce on the side is easier to see and to count.
• Say what the photo can’t show. Cooking oil, butter, dressing, sugar in your coffee, whole or skim milk. If your app lets you add a note, use it.
• Check the ingredients and amounts. If an app shows what it counted, read it. A missing oil or an oversized portion is easy to spot. If it only shows a total, you have nothing to check.
• Scan packaged food instead. A barcode or the nutrition label is usually closer than a photo of the food.
• Weigh calorie-dense foods now and then. Oil, nuts, nut butter, cheese, rice and pasta. A kitchen scale for a week trains your eye for the rest of the time.
• Be consistent and watch the trend. Your weight over a few weeks tells you more than any single meal. If it doesn’t match what your log says, look first at the things photos miss: oil, sauces and drinks.

How CalSimple handles accuracy

CalSimple wasn’t part of the NIH study, and we haven’t tested its accuracy in a study like this. Its photo estimates can miss in the same ways. What it does is show its work, so you can catch the misses the study describes.

• Every ingredient it counted, with the amount. Each entry lists what it counted and how much, not just a total. Hidden extras it counted, like cooking oil, butter or dressing, get their own line. If the oil isn’t there, you can see it wasn’t counted.
• A confidence badge. Each photo estimate shows high, medium or low confidence. When it’s low, the app says it’s a best guess and asks how much you had.
• Fix it in a sentence. Tap Fix results and describe what was different, like “half the rice” or “cooked in two tablespoons of olive oil”. Or edit any number directly.
• Barcode, label or words. For packaged food, scan the barcode or snap the nutrition label. No photo? Describe the meal in words.

Here’s how CalSimple works, step by step. You get 2 free AI scans a day, and manual entry, weight logging and charts are free and unlimited. If you want a tracker that stays out of the way, here’s what we mean by a simple calorie tracker. Using Cal AI now? See how the two compare on our Cal AI alternative page.

Common questions

Can AI count calories from a photo?

Yes, as an estimate. The app names each food, guesses how much of it there is, and works out the calories, protein, carbs and fat from that. It does best when it can see everything on the plate, and worst with hidden oil, sauces and mixed dishes.

Are AI calorie apps accurate enough to track weight loss?

They’re good for seeing patterns and staying aware of what you eat, as long as you check portions and fats. A nutrition researcher quoted by Healio suggested treating them as tools for building habits rather than precise measurements. To set a starting daily target, try the calorie calculator. None of this is medical advice. Talk to a doctor or dietitian before big changes to how you eat.

Which foods are hardest for photo calorie counters?

Anything where the calories are hidden: fried food, dishes cooked in oil or butter, creamy sauces, dressed salads, and mixed dishes like curries and casseroles. In the NIH follow-up, high-fat ketogenic meals gave the apps more trouble.

Keep reading

• Calorie calculator
• Cal AI alternative
• No-subscription calorie tracker
• Simple calorie tracker