Millions of people open a calorie tracking app every day, snap a photo of their lunch, and trust the number that appears on screen to guide their weight management decisions. It feels scientific: a photograph, an algorithm, a precise calorie count. But new research presented at the NUTRITION 2026 conference suggests that confidence may be misplaced. A study from researchers at the National Institutes of Health found that four popular AI-powered calorie tracking apps consistently underestimated the calorie content of standardized meals by roughly one-third, with errors reaching 250 to 345 calories per meal on average.
The findings, presented July 26 at the American Society for Nutrition annual meeting by Aaron Hengist, a postdoctoral visiting fellow at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), and Olivia Charles, a postbaccalaureate fellow at NIDDK, raise uncomfortable questions about the rapidly expanding universe of consumer-facing health AI. These tools operate almost entirely outside medical device regulation, their inner workings are proprietary secrets, and as the data shows, their errors are not random noise but systematic distortions that could quietly sabotage the goals of the very people who rely on them.
The study was conducted within a larger NIH Clinical Center nutrition trial comparing low-carbohydrate (ketogenic) diets with standard dietary patterns. That setting gave the researchers a rare advantage: total control over what went onto the plate. Meals were prepared in a metabolic kitchen, a research-grade food preparation facility where every ingredient is weighed to within 0.1 grams. The true calorie and macronutrient content of each meal was known with laboratory precision. Photographs of 102 such meals were then submitted to four leading apps: MyFitnessPal, LoseIt!, CalAI, and Appediet. A second phase added more than 200 additional meals to the dataset.
The results were striking across all four platforms. On average, the apps underestimated the caloric content of meals by 250 to 345 calories. Given that the meals in the study ranged from moderate to large in size, this represents an error of approximately 30 percent below the true value. The fat content was underestimated by roughly 30 grams per meal, a substantial gap. Carbohydrate estimates, by contrast, were notably more consistent with the ground-truth values.
The pattern of error was not uniform across app or meal type, and that unevenness tells its own story. High-fat meals, including the ketogenic meals central to the parent NIH trial, produced the largest discrepancies. The apps consistently underestimated fat content, and consequently total calories, for these meals. This is not a trivial corner case. Ketogenic and other high-fat dietary patterns have surged in popularity, and many of the people following them are precisely the users who lean most heavily on tracking tools to stay within their macronutrient targets.
There were also meaningful differences between apps. MyFitnessPal and LoseIt! showed better accuracy for higher-calorie meals, suggesting that some algorithms may perform adequately within certain ranges while degrading significantly at others. CalAI and Appediet, meanwhile, underestimated calorie counts more consistently across the board, regardless of meal size or composition.
None of this should be surprising in retrospect, the researchers note, and yet it is. The apps present themselves to users as measurement instruments, and measurement instruments are supposed to be accurate. But unlike a bathroom scale or a blood glucose monitor, these tools are not regulated as medical devices by the Food and Drug Administration. They are consumer software products, subject to no standard of accuracy, and their algorithms are black boxes whose training data and failure modes are unknown to the public and often to the scientific community as well.
This is the black box problem of consumer AI health tools. When a user photographs a plate of eggs, avocado, and bacon, the app does not recognize the food in the way a human nutritionist would. It matches visual features against a training dataset of labeled images, estimates portion sizes through photogrammetry or heuristics, and then cross-references those estimates with a nutrient database. Each step in that pipeline introduces error. The visual recognition may confuse similar-looking foods. The portion size estimate may be thrown off by the angle of the photograph or the depth of the bowl. The nutrient database entry may correspond to a generic version of the food that bears little resemblance to what is actually on the plate. And because the training data for these systems is proprietary, no outside researcher can audit where the pipeline breaks down.
What makes the NIH findings particularly concerning is the systematic nature of the error. If an app overestimated calories as often as it underestimated them, a user tracking their intake over a week might land close to the truth through averaging. But a tool that consistently underestimates by roughly 30 percent is not just imprecise. It is biased. A person relying on such an app to maintain a 2,000-calorie-per-day target could be consuming 2,600 calories in reality, every day, without any signal from the tool that something is wrong. Over the course of a week, that discrepancy amounts to more than 4,000 unaccounted calories, enough to completely offset the caloric deficit targeted by many weight loss regimens.
The implications extend beyond individual user confusion. Consumer health data from apps like these increasingly feeds into research studies, electronic health records, and even clinical recommendations. If the tools generating that data harbor systematic errors of this magnitude, then any analysis built on top of them inherits those distortions. Public health researchers who rely on app-based dietary recall data may be drawing conclusions about population-level eating patterns that are significantly skewed, particularly for fat intake and for people following low-carbohydrate diets.
The researchers are careful to note that this was a preliminary conference presentation and has not yet undergone peer review. The sample of 102 meals in the first phase, while carefully controlled, is not large enough to characterize every edge case across every app version and platform. The second phase of more than 200 additional meals will help clarify the scope and consistency of the problem. But preliminary or not, the finding that four major apps share a consistent, substantial, and directionally biased error pattern is a red flag that the research community and the public should take seriously.
For now, the takeaway for users is not necessarily to abandon these tools, but to recalibrate what they think they are getting. An AI calorie tracker is not a calorimeter. It is a convenient estimate generated by a proprietary algorithm trained on unknown data, operating without regulatory oversight, and demonstrably prone to underestimating the very nutrients that matter most to the people using it. Weight management is difficult enough without placing blind trust in a black box that reliably tells you your meal was smaller than it really was.
The NIH team plans to continue analyzing the expanded dataset and to explore whether newer versions of these apps show improvement. But as the market for AI health tools accelerates, the burden should not fall solely on academic researchers to test consumer products that make implied medical claims. The question the study surfaces is not whether these apps are sometimes wrong. Every measurement tool is. The question is whether the public deserves to know how wrong, in which direction, and under what conditions before they put their health data and their dietary decisions into a black box.