Why AI photo calorie counts come out low

What the researchers did

The study comes from the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), part of the NIH. Olivia Charles presented it at NUTRITION 2026, the annual meeting of the American Society for Nutrition, on 25 July 2026; all ten authors are from NIDDK, with Aaron Hengist listed last (abstract, ASN press release).

The meals came from a controlled feeding study at the NIH Clinical Center that compares a low-carb ketogenic diet with a standard one. They were planned from national food composition tables and cooked in a metabolic kitchen, where ingredients are weighed to a tenth of a gram, which makes the reference values unusually precise.

The researchers photographed 102 of these meals and ran the photos through the AI photo features of four apps: MyFitnessPal, Lose It!, Cal AI and Appediet. The meals averaged 918 kcal (from 592 to 1,211), with 60 g of carbohydrate, 61 g of fat and 36 g of protein.

What they found

Every app underestimated the energy in the meals:

All four differences P < 0.001. Source: the abstract (Charles, Hengist and colleagues, NIDDK).

• Fat: all four apps underestimated fat by about 30 g per meal.
• Carbohydrates: Lose It! and Cal AI underestimated them by about 14 g; for MyFitnessPal and Appediet the difference was not significant. All four estimated carbohydrates more consistently than the other nutrients.
• Meal size: MyFitnessPal and Lose It! were more accurate for higher-energy meals than for lighter ones.
• Overall: about 33% too low, similar, the authors note, to the underestimation previously reported when people record their own diet.

After the first analysis the team looked at more than 200 further meals. Early results suggest the apps may have more trouble with low-carb ketogenic meals, which often contain more fat (ScienceDaily).

Why a photo misses fat

The study measured how much the apps missed, not why. Aaron Hengist put the practical side this way: “People using a photo-based tracking app without adjusting the portions or entering the amounts of food should take the results with a grain of salt,” and “These apps tend to underestimate calories, especially from fats” (ASN press release).

Our reading, which the study did not test: oil in the pan, butter in a sauce and dressing on a salad are hard to see in a picture, and they add up quickly. One tablespoon of oil is 119 kcal, the value Flawless uses.

What Flawless does about it

Flawless has an AI photo scan too, and it is an estimate like any other. What it adds is a question. After a plate photo, if the AI judges that the dish is usually cooked in oil or butter, or served with a sauce or dressing whose amount the photo cannot show, such as fried, sautéed, roasted or stir-fried food, pasta with sauce or salad with dressing, Flawless asks: “Cooked in oil or with a sauce?”

You answer with one tap: Without oil, +1 tbsp oil, +2 tbsp oil, or With sauce or dressing. Each tablespoon adds 119 kcal and 13.5 g of fat straight away, without another AI request. “With sauce or dressing” asks the AI to add a typical portion of the sauce or dressing the dish is served with, which counts as one more scan from your daily allowance. Under the question there is always a field where you can tell the AI what else it missed.

It does not ask after every photo: not for raw fruit and vegetables, food boiled or steamed without fat, drinks, or when the fat is visible and already counted. Scans of a nutrition label and meals you describe in words do not get the question.

We have not measured how much closer this brings the number, so we do not claim it fixes the problem. It puts the fat question in front of you while you can still answer it.

How to log closer to the real number

• Treat a photo count as a first draft. Check the portion before you save it, which is what the researchers advise.
• Add the cooking fat. If the dish was fried or dressed, add the oil or the sauce, in Flawless or in any other app.
• Scan the barcode of packaged food. A barcode looks the product up instead of estimating it from a picture. In Flawless barcode scanning is free and has no daily limit.
• Weigh what you eat often. Entering the grams once takes the guessing out of the foods you log every week.
• Take extra care on keto. Early results suggest high-fat ketogenic meals are harder for photo apps.

What the study does not tell you

• It is not yet peer reviewed. The abstract was selected by an expert committee but has not been through full peer review, so the results count as preliminary until they appear in a journal.
• It tested four apps. Flawless and other apps were not part of it, so it says nothing about how they would do.
• The meals came from a research kitchen. Restaurant food and home cooking were not part of this test.

More guides

• A Cal AI alternative with free photo scans
• The best calorie and macro tracker apps in 2026
• A MyFitnessPal alternative with a real workout log
• Apps that track workouts and food together

Questions

Are AI photo calorie counters accurate?

Not very, according to a preliminary NIH study presented at NUTRITION 2026: the photo features of MyFitnessPal, Lose It!, Cal AI and Appediet underestimated 102 measured meals by 252 to 345 kcal on average, about a third. The study is not yet peer reviewed.

Why do photo calorie apps undercount?

In the study, the gap all four apps shared was fat: about 30 g per meal. The study did not test why; oil, butter and dressings are hard to see in a photo.

Which apps were tested?

MyFitnessPal, Lose It!, Cal AI and Appediet. Flawless was not tested.

Is the Flawless photo scan more accurate?

We don’t know, because it has not been measured, and it is an estimate like the others. After a plate photo of a dish usually cooked with fat, Flawless asks about oil or sauce so you can add it with one tap.

How can I log food more accurately from photos?

Check the portion, add cooking oil and sauces, scan the barcode of packaged food, and weigh the foods you eat often.