How Accurate Are AI Calorie Photo Apps? What the Research Actually Says

Error rates by logging method

What the 52-study review actually found

Shonkoff and colleagues reviewed 52 papers published between 2010 and 2023 that tested AI-based image dietary assessment against ground truth. Their headline numbers: average relative errors of 0.10% to 38.3% for calories and 0.09% to 33% for volume, with individual errors reaching as high as 79.6% on the worst single estimates.

That is an enormous spread, and the spread is the finding. Their conclusion was carefully worded:

• "Results suggest that AI methods are in line with — and have the potential to exceed — accuracy of human estimations." — Shonkoff et al., Annals of Medicine, 2023 (PMID 38060823)
• The review also flags a real gap in the evidence: only one of the 52 papers directly compared AI accuracy against humans alone. So "AI beats people" is not yet a settled result — it is a reasonable inference from separate bodies of work.

Where the error comes from: it is the portion, not the food

Identifying what is on the plate is largely a solved problem. Estimating how much is not, and that is where nearly all of the error lives.

A photograph flattens a three-dimensional object. Nothing in a single 2D image tells a model whether that is a shallow bowl or a deep one, whether the rice is packed or fluffed, or how much oil went into the pan before the photo was taken. Invisible ingredients are the hardest case of all — two visually identical plates of pasta can differ by 300 kcal depending on how much butter is in the sauce.

• Simple, single foods estimate well. The review found relative errors were lower for images of single or simple foods — an apple, a chicken breast, a bowl of rice.
• Mixed dishes estimate badly. Stews, casseroles, layered dishes and anything where components are hidden are where the 30%+ errors cluster.
• Added fats are close to invisible. No photo-based system reliably detects the oil, butter or cream already absorbed into a dish.
• Reference scale helps. Including a known object — a fork, a standard plate, your hand — in the frame measurably narrows portion error.

The comparison that actually matters

Photo estimation is usually judged against a food scale, and against a food scale it loses. But almost nobody weighs their food. The real-world alternative is typing an estimate into a database, and that alternative has been measured too — repeatedly, and unflatteringly.

In the 1992 New England Journal of Medicine study that reshaped how nutrition researchers read food diaries, ten diet-resistant subjects were tracked for 14 days using doubly labelled water, the gold standard for measuring true energy expenditure. They underreported their food intake by 47 ± 16% and overreported their physical activity by 51 ± 75% — while sincerely believing their logs were honest.

This is not a story about dishonesty. It is a story about memory and portion blindness: people forget the handful of nuts, and they misjudge what 100 g of pasta looks like. A 2024 Nature Food analysis of 6,497 doubly-labelled-water measurements found misreporting in 27.4% of records when the method was applied across large dietary datasets.

So the choice is not between a perfect method and a flawed one. It is between two flawed methods with different failure modes — and one of them takes four seconds.

Why consistent error hurts less than you think

If your photo log reads 15% low every single day, your absolute numbers are wrong but your trend is intact. Eat 300 kcal more on Saturday and the log still shows Saturday as higher. Adjust your portions down and the log still moves down.

That matters because for a weight goal you are not chasing an absolute number — you are chasing a direction and a rate of change, which you then calibrate against the scale over two to three weeks. A method with consistent bias supports that. A method you abandon after nine days does not, no matter how precise it was while you used it.

This is the strongest practical argument for photo logging, and it is worth stating plainly: the accuracy that matters is the accuracy of the method you actually keep using.

• Weigh and log manually if you need absolute precision for a medical or competitive reason. Nothing beats a scale.
• Photograph if the realistic alternative is not logging at all — which, for most people who have tried manual logging, it is.
• Either way, calibrate against your own bodyweight trend over 2–3 weeks rather than trusting any daily total.

Where Suu sits — including what it does worse

Suu does photo calorie analysis, returning calories, protein, carbohydrate and fat from a picture of your plate, and it can also take the entry by voice. It carries exactly the same portion-estimation limits described above; we have no reason to claim otherwise and no measured accuracy figure of our own to publish.

Two things Suu does differently are worth naming, and neither is about accuracy:

• The drink is not lost. When you log "a bowl of soup and a glass of ayran", Suu separates the solids from the liquids: the soup becomes calories and macros, and the ayran also counts toward your hydration. Calorie-only apps discard the second half of that sentence.
• The meal changes your water target. Protein and sodium in a logged meal add a digestion-water requirement to your daily goal, and a recorded workout raises it further.
• Honest limits: Suu's food database is smaller than MyFitnessPal's, there is no barcode scanning, and if you deliberately track micronutrients, Cronometer is the better tool. Three AI analyses per day are free; unlimited analysis and photo recognition are Premium.

Frequently asked questions

References

1. Shonkoff E. et al. (2023) — AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review. Annals of Medicine (PMID 38060823)
2. Lichtman S.W. et al. (1992) — Discrepancy between self-reported and actual caloric intake and exercise in obese subjects. N Engl J Med 327:1893–1898 (PMID 1454084)
3. Calorie estimation from pictures of food: crowdsourcing study. JMIR (PMC6246963)
4. Predictive equation derived from 6,497 doubly-labelled water measurements enables the detection of erroneous self-reported energy intake. Nature Food (2024)
5. Validity and accuracy of artificial intelligence-based dietary intake assessment methods: a systematic review (PMC12229984)

Related articles