How accurate are AI photo calorie counters? Useful, and usually low.

Short answer

Close enough to be useful, low enough to correct.

A photo calorie counter is good at the things a camera can see: what the food is, how much of the plate it covers, whether it is a salad or a schnitzel. It is weak at the things a camera can’t: how deep the bowl is, how much oil went into the pan, and what is hiding under the noodles.

That is why the errors are not random. They lean low, and they lean low on fat.

The 2026 test

What researchers at the NIH found.

• Who: researchers at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), part of the US National Institutes of Health.
• The meals: 102 meals from the metabolic kitchen at the NIH Clinical Center, where every ingredient was weighed to a tenth of a gram, then photographed.
• The apps: MyFitnessPal, Lose It!, Cal AI and Appediet.
• The result: all four underestimated, by about 250 to 345 calories a meal on average — roughly a third of the meal.
• The pattern: fat was missed most, by about 30 g a meal; carbohydrates were estimated most consistently. A follow-up of more than 200 more meals found very high-fat meals hardest of all.
• The caveat: this was a conference abstract presented in July 2026, not yet a peer-reviewed paper, so treat the figures as preliminary. Meals from a research kitchen are also not restaurant meals.

Why they miss

Five things a camera can’t see.

• Depth. A photo is flat. A wide shallow bowl of pho and a deep one look alike from above.
• Cooking fat. The oil in a stir-fry, the butter in mashed potato, the lard in a hawker plate — invisible, and the richest part of the dish.
• Sauce that soaked in. Curry gravy in the rice, dressing at the bottom of the salad.
• What’s underneath. Brisket under the noodles, a second patty under the bun, pearls at the bottom of the cup.
• Sugar in drinks. A sweet iced tea and an unsweetened one are the same colour.

Eating out

Why restaurant food is harder still.

Restaurant kitchens cook richer and serve bigger than home kitchens, and portions vary by cook and by night. Even restaurants’ own printed numbers wobble: in one Tufts study of 269 chain foods, 19% were off by more than 100 calories against the lab.

Independent restaurants usually print nothing at all, and their meals are large — a Boston study measured them at about 1,327 calories on average with sides. A photo estimate that runs a third low on a meal that size misses a lot. Our guide to counting calories when eating out shows how to adjust by hand.

Use it well

How to read a photo estimate.

• Treat it as a floor, not a ceiling — if the dish was glossy, fried or creamy, nudge it up.
• Check the portion it assumed against what you actually ate, especially on shared plates.
• Correct the dishes you eat often once, so the next estimate starts from your number.
• Compare against a typical restaurant number in the dish-by-dish guide when something looks off.
• Don’t chase precision: a rough number you keep is worth more than a perfect one you give up on.

How PlateDiary does it

A rough number that admits it’s rough.

In PlateDiary, calories are off unless you turn them on. With them on, each dish on the table gets one estimate with a tilde — “~640 kcal” — rounded to 10, or to 50 past 1,000. The app tells you that photo estimates usually run low.

You can Accept the estimate or Edit it, and a correction is remembered for that dish at that place. We don’t publish an accuracy percentage for PlateDiary, because we haven’t measured one in a controlled test — and we would rather say that than invent a number. More on how the calorie counter works.

Know the limits

When a photo isn’t enough.

If you manage a health condition, need exact numbers, or are working with a care team, a photo estimate is not the right tool on its own. Ask a doctor or registered dietitian what they want you to measure, and how.

In PlateDiary

A rough number for the dishes you save — only if you want one.

Sources

Where the facts come from