Are AI calorie counters accurate? What a 2026 NIH study found

Updated October 11, 2026: added studies of ChatGPT, Claude and Gemini on food photos, and the NIH study’s published abstract.

Snapping a photo of your plate and getting a calorie count in seconds feels a bit like magic. It also makes tracking far less tedious, because typing in every ingredient gets old fast. But convenience only helps if the number lands in the right neighborhood.

In July 2026, researchers at the National Institutes of Health shared what they described as the first direct, high-quality comparison of its kind. The short version: on their own, photo estimates came in low, mostly because of fat. The longer version is more useful, because the biggest errors are also the easiest ones for you to fix.

What the NIH study tested

The study comes from scientists at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) and was presented at NUTRITION 2026, the annual meeting of the American Society for Nutrition, according to the society’s announcement. Its abstract, by Olivia Charles and colleagues, is published in Current Developments in Nutrition (doi:10.1016/j.cdnut.2026.108738).

The test meals came from a larger NIH Clinical Center study comparing a low-carb ketogenic diet with a standard diet. They were prepared in a metabolic kitchen, where ingredients are weighed to the nearest 0.1 gram, so the researchers knew exactly what was on every plate. They ran standardized photos of 102 of those meals through the photo features of four apps: MyFitnessPal, Lose It!, Cal AI and Appediet.

One important caveat: this was a conference abstract. The organizers note that findings like these haven’t gone through the full peer review required for journal publication and should be treated as preliminary.

What the researchers found

All four apps underestimated calories, by about 250 to 345 calories per meal on average, or roughly a third. Healio reported the average for each app:

A few other findings stand out:

• Fat was the big miss. All four apps underestimated fat by about 30 grams per meal.
• Carbs were steadier. Carbohydrate estimates were more consistent than those for other macronutrients, although Lose It! and Cal AI still underestimated carbs by about 14 grams.
• Meal size mattered for some apps. MyFitnessPal and Lose It! estimated higher-calorie meals more accurately than lower-calorie ones.
• High-fat meals looked hardest. In a follow-up of more than 200 additional meals, early results suggested the apps struggled more with ketogenic meals, likely because their higher fat content was consistently underestimated.

Apps update often, so these results describe the versions tested for this study, not necessarily how each app performs today. Calorix wasn’t part of the study, and we don’t assume a photo-only estimate from any app, ours included, would escape the same problems.

The advice from Aaron Hengist, one of the researchers, was plain: people using a photo-based app “without adjusting the portions or entering the amounts of food should take the results with a grain of salt.”

This isn’t only an AI problem

It helps to compare the apps with the usual alternative, which is people estimating for themselves. The researchers noted that a one-third underestimate is similar to what has been reported for self-reported diets.

That fits decades of research. In the OPEN study, which checked the self-reports of 484 adults against objective biomarkers, 24-hour recalls underreported energy intake by about 12 to 20 percent, and food frequency questionnaires by roughly a third. A 2019 systematic review of 59 studies using doubly labeled water found that most reported significant underreporting, including studies of technology-based methods.

Photos don’t fix that automatically. A 2020 meta-analysis found that image-based dietary assessment also tended to underreport energy and concluded that, like traditional methods, it carries serious measurement error. A 2023 systematic review of fully automated AI methods found average calorie errors ranging from under 1 percent to about 38 percent across studies, with smaller errors for images of single, simple foods.

There is a hopeful detail, though. In the remote food photography method, where researchers estimated intake from photos of each meal and its leftovers, estimates came within about 5 to 7 percent of weighed intake (still slightly low). Photos can work well when a person reviews them carefully, and that review is exactly what the NIH test left out.

The commentary that accompanied the Healio report made one more point worth remembering: a consistent underestimate matters more than random error. Random errors partly cancel out over a week; a bias in one direction keeps adding up.

What about ChatGPT and other AI chatbots?

General AI models make the same kind of errors. In a 2025 study of 52 standardized food photos, ChatGPT-4o and Claude 3.5 Sonnet missed the calories by 35.8% on average and Gemini 1.5 Pro by 64.2%, with all three under-counting more as portions grew. In another 2025 study, ChatGPT-4 identified 93% of the foods in 114 meal photos but under-estimated the weight of medium and large meals. And the details you give matter: told the cooking fat and the ingredient amounts, ChatGPT-5’s average calorie error on 195 dishes fell from 30.5% to 13.9% (2025 study). More in Can ChatGPT count calories from a photo?, and a look at each photo app’s accuracy claims in apps that count calories from a photo.

Why photo estimates miss

Hidden fats

Oil soaks into rice and vegetables. Butter melts into sauces. Dressing coats every leaf. A camera sees none of it, and fat is calorie-dense: USDA FoodData Central lists one tablespoon of olive oil at 13.5 grams of fat and 119 calories. The roughly 30-gram fat gap in the NIH study is a little more than the fat in two tablespoons of oil.

Portion depth

A photo is flat. From directly above, a shallow layer of pasta and a deep bowl of it can look almost the same. In a 2025 study, adults who matched real rice portions to photos tended to underestimate them in overhead shots, because the height of rice in a bowl is hard to see from above. Photos taken at a 45-degree angle did best for solid foods.

Mixed dishes

Curries, casseroles, burritos and sandwiches hide ingredients inside or under each other. The AI has to guess the recipe as well as the amount. Fittingly, the 2023 review found errors were smaller for images of single, simple foods.

Lighting and angle

Image researchers note that changes in illumination and viewpoint change how foods look, which makes them harder to identify. Dim restaurant lighting, strong shadows and unusual angles all add uncertainty.

How to get better estimates

You don’t need to weigh everything to do much better than a raw photo estimate. These habits target the errors above:

• Add a note when you scan. Tell the app what it can’t see: “cooked in butter”, “extra dressing”, “half portion”.
• Check portions, not just the total. Look at the grams for each ingredient and fix anything that looks off, starting with the calorie-dense items.
• Log fats as their own line. Add the oil, butter, mayo or dressing as a separate ingredient so it doesn’t vanish into the dish.
• Use barcodes and labels for packaged food. The label gives you the manufacturer’s numbers, so there’s nothing to guess from a photo.
• Weigh occasionally. A week of weighing your usual rice scoop, cereal bowl or pour of oil recalibrates your eye. Precisely weighed meals were the NIH researchers’ reference point for a reason.
• Take a better photo. Good light, the whole plate in frame and a 45-degree angle all help. Our guide to estimating portions from a photo covers this in detail.
• Expect more fat when someone else cooks. It’s usually the biggest unknown in a restaurant meal. See how to log restaurant meals.

How Calorix is designed around the problem

We built the Calorix AI scanner on the assumption that every photo estimate is a first draft. Instead of a single number, a scan opens a review card that lists each ingredient with its portion, grams, calories and a confidence badge, so you can see where the estimate is least sure.

From there you can:

• edit names, portions and grams, or add and remove ingredients
• use the portion slider (½, 1, 1½ or 2) when you ate more or less than the photo shows
• tap Fix with AI and type or say a correction, such as “no rice, add 2 eggs”, then review the proposed changes before accepting them
• add a note like “cooked in butter” and re-analyze from the stored photo
• override any value by hand, with the AI’s original kept so you can revert to it

Totals are always recalculated from the ingredients, so a correction to one item flows straight through to your daily calories and macros. For packaged food, barcode and nutrition-label scanning skip the guesswork.

To be clear about what we’re not claiming: we haven’t shown that Calorix is more accurate than the apps in this study. What the design does is make the estimate transparent and quick to correct, which is the step the NIH researchers pointed to.

The bottom line

AI photo logging is a fast way to get a reasonable first draft, but on its own it tends to run low, especially for oily, buttery and mixed dishes. Treat every photo estimate as a starting point: check the portions, add the fats you know are there, and use barcodes for anything with a label.

Small corrections add up. If what you eat runs a few hundred calories above what your log says, the plan you built won’t match reality, and it’s easy to blame yourself instead of the estimate. A realistic log, paired with a realistic target from our TDEE calculator, gives you numbers you can actually work with.