3 min read
If you have spent 2026 photographing your lunch so an AI app can tell you how many calories are in it, here is some bad news wrapped in slightly liberating news: the app is probably wrong, and it is wrong in the direction you would least prefer. A new study has found that popular photo-based calorie-tracking apps underestimate the calories in your meal by about a third.
What the study found
Researchers at the US National Institutes of Health ran a properly careful test. They took standardised photos of 102 meals, prepared as part of a controlled diet study where the true nutritional content was known precisely, and fed them through several popular AI-powered apps, including big names like MyFitnessPal and LoseIt. The apps use image recognition to guess what is on the plate and how much of it there is, then look up the nutrition. The result: the apps underestimated energy by roughly 250 to 345 calories per meal and undercounted fat by around 30 grams. That is not a rounding error. That is an entire snack the app politely forgot to mention.
Why it gets it so wrong
The reason is obvious once you think about it. A photo is flat, and calories hide in three dimensions and in the invisibles. The AI cannot see the tablespoon of oil the vegetables were fried in, the butter melted into the mash, or how deep the bowl really is. Fat is calorie-dense and frequently invisible in a picture, which is exactly why the apps did worst on high-fat, low-carbohydrate meals, the sort a lot of keto and low-carb dieters eat precisely because they are trying to count. The people most likely to trust the number are the ones getting the least accurate one.
What to do with this
The point is not to count harder or panic over a few hundred calories, which is a fast track to a joyless relationship with food. The point is to keep the technology’s confidence in perspective. These apps hand you a specific, official-looking number, and that precision is largely an illusion generated by a model making an educated guess from a picture. If you find food tracking genuinely useful, treat the figure as a rough ballpark rather than gospel, and assume the real total is meaningfully higher, especially for anything cooked in fat. And if a photo of your dinner is what makes you anxious, the healthiest thing the research quietly points to is putting the app down altogether.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and with the horseshit filter set to maximum. Free.
Did you know: estimating portion sizes is one of the hardest problems in nutrition science, and even trained dietitians routinely misjudge them from photographs. Asking an AI to do it from a single snapshot is asking it to solve a problem the experts have not cracked either. The confident number on your screen is best read as a guess wearing a lab coat.