Guide

How accurate are AI calorie counters?

Less accurate than their marketing suggests. Research presented at NUTRITION 2026 found four popular apps underestimated meals by 250 to 345 calories on average, and fat by around 30 grams. Photo estimation is good at naming food and weak at judging portions — which is exactly where the calories hide.

Written by the Calybite team · Last updated 24 August 2026

What the research actually found

Researchers at the National Institute of Diabetes and Digestive and Kidney Diseases, part of the US National Institutes of Health, tested four apps — MyFitnessPal, LoseIt!, Cal AI and Appediet — against known meal compositions. Across the set, estimated calorie totals came in roughly 250 to 345 calories too low per meal, and fat was underestimated by about 30 grams. Carbohydrate estimates were the most consistent; high-fat ketogenic meals gave the systems the most trouble.

An honest caveat about this finding. It was presented as a conference abstract at NUTRITION 2026 and has not yet completed peer review. The authors describe the results as preliminary. We are citing it because it is the best independent evidence currently available — not because a conference abstract settles the question. Treat the direction of the finding as more reliable than the exact figures.

Why portions are the hard part

Photo-based estimation splits into two jobs, and they are not equally difficult. Naming what is on the plate is largely solved — modern vision models identify chickpeas, rice, a boiled egg and a chicken thigh reliably. Judging how much of each is there is not.

A photograph is two-dimensional. A wide flat steak and a thick narrow one look nearly identical from above, and the difference can be two hundred calories. Separate work on image-based dietary assessment found portion-size estimation reliable for only about 39% of dishes tested. That is the number that matters, because portion error propagates directly into calorie error.

Then there is what a camera simply cannot see: the oil the kitchen cooked in, the butter under the sauce, the sugar in the marinade. These are invisible, calorie-dense, and they run in one direction — which is why the errors are underestimates rather than a spread around the truth.

Be sceptical of vendor accuracy claims

Almost every app in this category publishes an impressive-sounding accuracy figure — 90%, 97.4%, plus or minus 1.9%. Nearly all of them are self-scored, on undisclosed test sets, by the company that wins the comparison. Cal AI’s widely quoted 90% figure has been publicly challenged for having no citation attached.

A useful rule: if a number about a product comes from the company selling it, and no method is published, treat it as marketing. That applies to us too — which is why there is no Calybite accuracy percentage anywhere on this site. We do not have an independent one, so we are not going to invent one.

So should you use one?

Yes, with your eyes open — because the comparison that matters is not photo-logging versus perfect measurement. It is photo-logging versus not logging at all.

Weighing everything on a kitchen scale is more accurate, and most people stop doing it within a fortnight. An approach that is 80% accurate and survives six months beats one that is 95% accurate and survives three weeks, because a calorie deficit only exists if you keep track of it. The failure mode of manual logging is abandonment, and abandonment has an accuracy of zero.

How to get more out of it

Three habits close most of the gap, whichever app you use:

Correct the portion, not the dish. The name is usually right; the amount is where the error lives. If the estimate says 150 g of rice and it was closer to 250 g, fix that one field and you have removed most of the error in the meal.

Add the cooking fat yourself. A tablespoon of oil is about 120 calories and no camera will find it. If you cooked it or the restaurant did, add it manually.

Judge on weekly trends, not single meals. Individual estimates scatter; the average over a week is far more stable and is the only thing your weight responds to anyway.

Where Calybite sits on this

Calybite, the AI food-tracking app, is built around the assumption that the estimate needs checking. Recognition is component-level, so you see chickpeas, rice and egg rather than “a bowl”. Portions are matched against verified nutrition entries rather than generated freehand. And nothing is written to your diary until you tap Save — the review screen exists specifically because portion estimation is the weak step, and a one-tap correction is faster than a re-scan.

That does not make it immune to the findings above. It makes the weak step visible and editable instead of silent, which is a different and more honest claim than a percentage. If you want the mechanics, what actually happens when you scan a plate walks through all three steps.

Related reading: how to read your calorie number — worth knowing what target you are measuring against before worrying about measurement error. And if you want the target itself, work out your daily calories.

Sources

Hengist A, Charles O, et al. Abstract presented at NUTRITION 2026, American Society for Nutrition, July 2026. Reported by ScienceDaily. Preliminary — not yet peer-reviewed.
Crowdsourcing and image-based dietary assessment studies on portion-size reliability, including Calorie estimation from pictures of food (PMC6246963).
Atwater energy factors — protein 4 kcal/g, carbohydrate 4 kcal/g, fat 9 kcal/g — why a 30 g fat error is worth roughly 270 calories.