A photo calorie app can look precise and still be wrong by a third of the meal. In July 2026, researchers at the U.S. National Institutes of Health presented that finding at the American Society for Nutrition's annual meeting: four popular photo-based AI calorie apps underestimated energy by about 250 to 345 calories per meal, and fat by about 30 grams.
That is not a rounding error. If you scan lunch and dinner without editing, you can "eat at maintenance" on the screen and still be in a surplus. This article explains what the study actually measured, why fat is the usual miss, and how to log so the number is usable - without throwing the scanner away.
Disclosure: SmartEat is our product. We were not one of the four apps in this test. The researchers' advice still matches how we think scanning should work: treat the photo as a draft, then fix portions and cooking fats.
What the 2026 study found
The team, including Aaron Hengist at NIH's National Institute of Diabetes and Digestive and Kidney Diseases, used 102 meals prepared in a controlled metabolic kitchen. Every ingredient was known. They photographed those plates and ran the same pictures through four consumer apps:
Largest average miss in this test. Photo-first logging with no barcode fallback in the study workflow.
Photo estimate ran low on energy and fat. Better on higher-calorie plates than lighter ones.
Meal Scan / photo AI was low here even though the app's barcode database is industry-scale.
Smallest miss of the four - still about a quarter of the meal's calories missing on average.
Across all four, calories and fat were about one third too low. Carbohydrates were estimated more consistently than fat. High-fat and ketogenic-style plates were the hardest. Hengist's takeaway, in the NIH press summary: if you do not adjust portions or amounts, take the result with a grain of salt - what you ate is likely higher than what the app shows.
Read the coverage: ScienceDaily, EurekAlert (ASN), Healio. The work was presented at NUTRITION 2026 (Charles O. et al.) and is preliminary until a full peer-reviewed paper is out - the direction of the error is still the useful part for anyone logging with a camera.
Why the camera misses fat
Identifying "chicken and rice" is the easy half. Estimating how much oil hit the pan from a flat photo is the hard half. A tablespoon of olive oil is about 120 calories and almost invisible once it soaks into food. Dressing, butter, cheese, creamy sauces, and fried coatings do the same job.
That is why this study's fat miss (~30 g, roughly 270 calories of fat alone) lines up with what dietitians see in real diaries. The AI is not "lying." It cannot see volume, hidden lipids, or whether the salmon was grilled or pan-fried in oil.
A scan is still useful - if you treat it as a draft
Throwing out photo logging would be the wrong lesson. People quit tracking when logging is slow. The study's risk is the opposite: logging that is fast and systematically low. Use the camera for speed, then spend 15 seconds on the three things AI is bad at.
- Confirm the items - remove duplicates, fix misreads (yogurt vs ice cream, soup vs stew).
- Fix the portion - if it looks like 1.5 servings, say so. Do not leave the default "1 plate."
- Add the invisible calories - oil, butter, dressing, cream in coffee, cooking spray used like oil.
If a food has a barcode, scan the pack instead of photographing the bowl. Label data beats a camera guess for yogurt, bars, sauces, and oils. For the method-by-method breakdown, see Photo vs barcode vs voice calorie logging.
Five checks after every photo scan
- Did I cook this in oil or butter? Add it.
- Is there sauce or dressing on the plate or on the side that I used?
- Is the protein skin-on, breaded, or fried?
- Does the portion match what I actually ate - including seconds?
- Would a barcode or a saved "recent" be more accurate for any item?
Do this for a week on meals you already eat. You will see the same pattern the NIH kitchen saw: fat and energy jump when you stop trusting the first number.
What this means if you are trying to lose weight
A 300-calorie miss at dinner, six nights a week, is about 1,800 calories you never logged. That can look like a stall, "broken metabolism," or a reason to quit the app. Often the log was optimistic, not the body.
Photo AI is still better than not logging. It is not a food scale. For tight deficits, medical diets, or GLP-1 eating where every bite counts, combine:
- Photo for mixed plates you will actually log
- Barcode for anything with a label
- A few weighed repeat foods (oats, oil, chicken) saved as recents
- A human review if the scale and the diary disagree for two weeks
Related: Food diary mistakes that slow progress and when to see a dietitian vs using an app.
How SmartEat is built for this (not a claim we "won the study")
SmartEat was not in the four-app test. We do not claim a magic accuracy score from this paper. What we built is the workflow the authors pointed at:
- Edit after every scan - names, grams, and extras before the meal hits your diary
- More than photo - meal, menu, fridge, and barcode so you are not stuck guessing a labeled oil from a picture
- Same diary - scans sit next to manual logs, macros, and plans, so a bad estimate is visible in the week, not a one-off screenshot
- Optional dietitian - when the camera and the scale still disagree
For which scanner to use on which meal, see Best AI meal scanner in 2026. This post is only about trusting the calorie number.
Bottom line
In 2026, photo AI is fast and systematically low on calories and fat - about a third low in this NIH kitchen test. Keep scanning. Stop treating the first number as weighed food. Edit portions, add oil, use barcodes, and look at the week, not one pretty plate.
Related reading:
