Your Photo Calorie Tracker Has Been Lying to You (By About 345 Calories)

Somewhere out there, a plate of chicken thigh, roasted potatoes, and a very confident drizzle of olive oil got photographed, uploaded, and told it was 340 calories. It was, in reality, closer to 650. The app did not apologize. The app never apologizes. The app just sits there, smug, glowing green because you're "under budget," while your body quietly files a complaint that will show up on the scale in about three weeks.

This isn't a hypothetical. It's the actual finding of a study presented July 25 at NUTRITION 2026, the flagship annual meeting of the American Society for Nutrition, in National Harbor, Maryland. Researchers Olivia Charles and Aaron Hengist, working with the National Institute of Diabetes and Digestive and Kidney Diseases, fed four widely used photo-based calorie apps — MyFitnessPal, Lose It!, Cal AI, and Appediet — 102 precisely, laboratory-measured meals. Every single app underestimated the calories. Every single one underestimated the fat, too, by roughly 30 grams a meal. The apps were off by an average of 250 to 345 calories per meal. That's not a rounding error. That's a second breakfast.

The pitch was too good, which should have been our first clue

"Just take a photo" has been the single most seductive sentence in diet tech for two years running. No searching a database. No weighing anything. No admitting, out loud, that you added a second scoop of rice. You point your phone, the AI squints at your lunch, and a number appears — clean, confident, and apparently made up. It turns out an image is a genuinely hard thing to convert into a calorie count, because a photo can't see what's underneath the surface: the tablespoon of butter melted into the mashed potatoes, the true depth of the bowl, whether that's olive oil or an optical illusion. The study found the apps were most accurate on carbs — bread, rice, pasta are visually legible — and worst on fat, which is exactly the macronutrient that's easiest to hide and hardest to estimate by sight. Keto and low-carb meals, being higher in fat, got hit hardest of all.

So what was actually happening in your logs

• Your "1,600-calorie day" was probably closer to 1,900–2,000. A per-meal miss of 250–345 calories compounds fast across three meals.
• Your fat intake was underestimated by roughly 30 grams a meal — about 270 extra calories nobody counted, hiding in dressing, oil, and marbling the camera couldn't weigh.
• Carbs were the app's best subject — visually obvious, structurally simple, hard to hide. If your diet is mostly rice and toast, the miss is smaller. If it's stir-fry and steak, the miss is bigger.
• Plateaus that felt mysterious weren't mysterious. "I'm eating 1,500 and not losing weight" is a much less confusing sentence once you know the app was quietly rounding down by a few hundred calories a day, every day, for months.

The apps aren't lazy, they're guessing — because a photo is a guess

This isn't really an indictment of any one app's engineering. It's a limits-of-physics problem. A 2D photo of a 3D plate is missing depth, missing what's mixed in, missing the oil that soaked into the rice before the photo was even taken. The researchers' own recommendation was a hybrid approach — combining the photo with some form of direct dietary input to correct for what the camera can't see. Which is, funnily enough, a pretty good description of just telling the app what's actually in the bowl instead of asking it to reverse-engineer your dinner from a JPEG.

This doesn't mean tracking is pointless — it means the method matters

The instinct after reading a study like this is to throw the whole idea of tracking out the window: if the apps are wrong, why bother? That's the wrong lesson. Every method of estimating calories has error bars — nutrition labels, restaurant menus, and food scales all carry some slop too. The actual lesson is narrower and more useful: photo-based estimation specifically struggles with fat and hidden ingredients, so if your tracking method can't capture "and then I added butter," your numbers are structurally biased low, not randomly off. A method that lets you say the whole truth about a meal — sauce, oil, the handful of nuts you ate standing at the counter — will systematically beat a method that can only see what's visible from above.

What to actually do about it

If a photo-based app has been your whole system, don't panic-recalculate every past week — just adjust going forward. Mentally pad any photo-estimated meal by 15–20%, especially anything with a visible sauce, dressing, or fried component. Better: for meals you cooked yourself, where you actually know the ingredients, say them instead of photographing them — you have information the camera never had access to. And if you've been stuck on a plateau you couldn't explain, or you're chasing a specific protein target, this is very likely part of why the math wasn't adding up. The app wasn't broken. It just couldn't see the butter.

Frequently asked questions