AI Calorie Apps Underestimate Meals by up to 345 Calories: What the 2026 NIH Study Found

---

title: "AI Calorie Apps Underestimate Meals by up to 345 Calories: What the 2026 NIH Study Found" description: "NIH researchers photographed 102 meals from a metabolic kitchen and logged them in four popular photo-based calorie apps. Every app came in low. Here is what the study actually tested, and what it means for how you log." publishedAt: "2026-08-25" updatedAt: "2026-08-25" author: "Inlab Products" category: "Accuracy" tags: ["AI calorie tracker accuracy", "photo calorie tracking", "calorie app study", "under-logging", "food photo recognition"] keyTakeaways:

• "Researchers at the NIH's NIDDK tested four widely used photo-based calorie apps against 102 meals of known composition. The apps underestimated by roughly 250 to 345 calories per meal on average."
• "Fat was the worst-estimated macro, off by about 30 grams. Carbohydrate estimates were the most consistent across all four apps."
• "High-fat and ketogenic meals were the hardest for the apps, which follows directly from fat being the macro a camera cannot see."
• "The critical detail most headlines skipped: meals were logged from the photo alone, with no portion adjustment. That is the worst case for any photo tracker, not normal use."
• "This is a conference abstract, not a peer-reviewed paper. Treat the direction of the finding as solid and the exact numbers as provisional." faq:
• question: "Are AI calorie tracking apps accurate?" answer: "They are directionally useful and consistently low. In the 2026 NIDDK analysis, four popular photo-based apps underestimated meals of known composition by roughly 250 to 345 calories on average, with fat under by about 30 grams. That said, the meals were logged from photos alone without any portion correction. An app you review and correct before saving behaves very differently from one you let auto-log."
• question: "Why do photo calorie apps underestimate calories?" answer: "A photograph carries no information about density or hidden fat. Oil absorbed into a stir-fry, butter melted into a sauce, the depth of a bowl you are looking down into, and dressing already tossed through a salad are all invisible to a camera. Fat is nine calories per gram and the least visible macro, which is why fat errors dominate the total calorie error."
• question: "Should I stop using a photo calorie tracker?" answer: "No, but stop treating the first number as the answer. Use the photo to get a fast draft, then correct the parts you know the camera missed: cooking oil, butter, dressings, and portion size. A corrected photo log takes about twenty seconds and lands far closer than either an uncorrected photo or a meal you never logged at all."
• question: "Which foods are hardest for AI calorie apps to estimate?" answer: "Mixed dishes where ingredients are no longer distinguishable, anything cooked in a fat you cannot see, soups and stews where depth hides volume, and high-fat or ketogenic meals. Plated whole foods with a clear size reference in frame are the easiest case and the one where these apps do well."
• question: "Was the NIH calorie app study peer reviewed?" answer: "Not yet. The findings were presented as an abstract at NUTRITION 2026, the American Society for Nutrition's annual meeting, held July 25 to 28 in National Harbor, Maryland. Abstracts presented there have not generally completed the peer review required for journal publication, so the results are preliminary."

---

If you have used a photo-based calorie tracker and quietly wondered whether the number was too good to be true, there is now research pointing the same way. Four widely used photo-based calorie apps underestimated meals by roughly 250 to 345 calories each.

That is the headline. The detail underneath it is more useful than the headline, and it changes what you should do about it.

TL;DR

Researchers at the NIH photographed 102 meals of exactly known composition and logged each one in four popular photo-based apps, without adjusting any portions. Every app came in low, by 250 to 345 calories and about 30 grams of fat. Fat was the problem macro. High-fat and keto meals were worst. The fix is not to abandon photo logging, it is to stop letting the first estimate stand. Review it, add the oil and butter the camera could not see, and correct the portion. That takes seconds and removes most of the error.

What the researchers actually did

Aaron Hengist and Olivia Charles at the National Institute of Diabetes and Digestive and Kidney Diseases, part of the NIH, built the test around meals whose composition was already known down to the gram. The meals came out of a controlled metabolic kitchen, the kind used for inpatient feeding studies where every ingredient is weighed.

They photographed 102 of those meals and logged each photo in four widely used photo-based calorie apps. Then they compared what the apps reported against the kitchen's own numbers. A later analysis extended the work past 200 meals.

The design matters. Most accuracy claims in this category, including our own 20-meal benchmark, compare an app estimate to a kitchen scale reading. A metabolic kitchen is a stricter reference than that.

The numbers

Feature	What the study found
Average calorie underestimate per meal	250 to 345 kcal
Average fat underestimate per meal	about 30 g
Most consistently estimated macro	Carbohydrate
Least consistently estimated macro	Fat
Meals estimated best	Higher-calorie meals
Meals estimated worst	Light meals, and high-fat or keto meals
Apps that underestimated	All four

Two things stand out.

The first is that all four apps missed in the same direction. That is not four separate bugs, it is a structural property of estimating calories from an image.

The second is that the error scales with how much fat a meal is hiding. Roughly 30 grams of fat is about 270 calories, which accounts for most of the gap on its own. The calorie error is largely a fat error wearing a different label.

Why a photo cannot see fat

Think about what a camera actually captures. Shape, colour, surface texture, and approximate area. That is enough to recognise chicken, rice, and broccoli. It is not enough to answer the question that determines the calorie count, which is how much fat went into all three.

The specific blind spots:

• Absorbed oil. A stir-fry cooked in two tablespoons of oil looks identical to one cooked in a dry pan. That is roughly 240 calories that leave no visual trace.
• Butter and fats melted into food. Butter in mashed potato, cream in a soup, ghee in a curry base.
• Dressing already tossed through. A salad with the dressing on it looks like a salad.
• Depth. Looking down into a bowl tells you the surface area of the food, not how far down it goes.
• Density. Two visually identical scoops of rice can differ by a third in weight depending on how tightly packed they are.

Fat is nine calories per gram against four for protein and carbohydrate, so every gram the camera misses costs more than twice as much. It is also the macro most likely to be invisible. Those two facts together explain the whole result.

It also explains the keto finding. A ketogenic meal is deliberately built to carry most of its energy as fat, which is exactly the energy a photograph cannot see. The apps struggled most with the meals where the invisible macro dominated.

The caveat the headlines skipped

Here is the sentence from the research team that most coverage left out, and it is the one that should change how you read the numbers:

"People using a photo-based tracking app without adjusting the portions or entering amounts of food should take the results with a grain of salt."

The meals were logged from the photo alone. No portion correction, no adding the oil, no editing the entry. That is a deliberate and defensible study design, because it isolates what the model does on its own. It is also the least accurate way anyone actually uses these apps, or should.

So read the result as a ceiling on the error, not a description of your Tuesday. An app you correct before saving is a different instrument from one you let file the estimate unseen.

One more caveat, and it is a real one. This was presented as an abstract at NUTRITION 2026, the American Society for Nutrition's annual meeting in National Harbor, Maryland, in late July 2026. Conference abstracts have not generally been through journal peer review. The direction of the finding is consistent with everything else known about estimating food energy from images. The exact figures should be treated as provisional until the full paper lands.

What to do about it

A 300-calorie daily error is not a rounding difference. It is larger than the deficit most people are aiming for, which means it can make a genuine deficit look like a plateau and send you hunting for a metabolic explanation that does not exist. We wrote the diagnostic order for that in calorie deficit but not losing weight.

Five corrections, in order of how much they buy you:

1. Review every estimate before it saves. This is the whole game. An estimate you never looked at is an estimate you never corrected, and the study is a measurement of exactly that scenario.
2. Add the fat you know is there. If you cooked in oil, log the oil. If there is butter in it, log the butter. You do not need to see it to know it went in. This single habit closes most of the gap the study measured.
3. Correct the portion, not the food. The recognition is usually right about what the food is. The quantity is where it drifts. Adjust grams or servings and leave the identification alone.
4. Weigh the few things that swing hardest. You do not need a scale for a chicken breast. You do need one for oil, nut butter, cheese, granola, and anything else where a small visual difference is a large calorie difference. Full list in how to track calories accurately.
5. Judge the app on your trend line. If your logs say 1,800 and your weight is flat for three weeks, the logs are wrong before the formula is. Calibrate against what your body does, not against what one meal estimate says.

The honest position on photo logging

It would be easy to read this study as a case against AI calorie tracking. I do not think that holds up, for a reason the study was not designed to measure.

The alternative to an imperfect log is usually no log. Self-reported food intake has been found to be under-reported by 20 to 40 percent across decades of research, and that is people typing entries by hand, carefully, in a study they know they are part of. A 300-calorie error from a photo you actually took is not obviously worse than a 600-calorie error from a manual entry you eyeballed, or than the week you stopped logging entirely because searching a database for every ingredient got old.

The right conclusion is narrower and more useful. Photo logging is a fast first draft, not a final answer. Treat the estimate as a starting point that needs one round of correction, and the speed advantage survives while most of the error does not.

Related reading

• How to Track Calories From a Photo covers the photo setup that gives any app the best chance, plus our own benchmark.
• How to Track Calories Accurately is the practical version: what to weigh, what to estimate, and the four hidden-calorie categories.
• The Complete Guide to AI Calorie Tracking explains how the underlying computer vision works and where it breaks down.
• Calorie Deficit but Not Losing Weight? walks the diagnostic order when the numbers and the scale disagree.

Sources

1. Charles O, Hengist A et al. (2026). "Accuracy of photo-based dietary tracking applications." Presented at NUTRITION 2026, American Society for Nutrition, National Harbor, MD, July 25 to 28, 2026. Press release: https://www.eurekalert.org/news-releases/1136415
2. American Society for Nutrition, NUTRITION 2026 meeting coverage. https://www.sciencedaily.com/releases/2026/07/260726015237.htm
3. HealthDay. "How Accurate Are Photo-Based Calorie Apps? 4 Are Put To The Test." https://www.healthday.com/health-news/nutrition/how-accurate-are-photo-based-calorie-apps-4-are-put-to-the-test
4. Hall KD, Guo J (2017). "Obesity Energetics: Body Weight Regulation and the Effects of Diet Composition." Gastroenterology. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5568065/
5. USDA FoodData Central. https://fdc.nal.usda.gov/