Millions of people snap photos of their plates each day, trusting AI-powered apps to count the calories for them. A new evaluation of four popular trackers found that every single one fell short, underestimating prepared meals by roughly 250 to 345 calories on average and missing about 30 grams of fat per meal. The tested apps, MyFitnessPal, LoseIt!, CalAI, and Appedi, were evaluated against 102 precisely prepared meals, and none came close to accurate totals. For anyone relying on these tools to manage weight, diabetes, or heart health, the gap between what the app reports and what is actually on the plate could quietly derail their goals.
Why a 345-calorie gap changes daily diet decisions
A shortfall of 250 to 345 calories per meal is not a minor rounding error. Across three meals a day, that could mean 750 to more than 1,000 untracked calories, enough to erase a calorie deficit entirely or push someone well past their daily target without any visible warning. The problem is especially acute for people following structured eating plans or managing chronic conditions where precise intake matters.
The evaluation, presented at the NUTRITION 2026 conference, used standardized photographs of 102 meals that had been weighed and prepared to exact specifications. This controlled setup removed the usual excuses about bad lighting or awkward camera angles. The apps still missed the mark. Fat content was underestimated by roughly 30 grams on average across all four platforms, a gap that carries outsized caloric weight because each gram of fat contains nine calories.
One plausible explanation is that the AI models behind these apps were trained largely on images of single-item plated dishes, the kind of clean, well-lit food photos common in restaurant databases and stock libraries. When a meal contains multiple mixed ingredients, layered sauces, or hidden fats like butter melted into grains, the model struggles to decompose the image into accurate components. The more complex the plate, the larger the blind spot. This pattern showed up clearly in the test results: ketogenic and other high-fat dishes proved particularly problematic for all four apps, suggesting that calorie-dense ingredients are the hardest for current vision models to detect and quantify from a photograph alone.
102 controlled meals and a separate free-living study confirm the pattern
The strength of the NUTRITION 2026 findings lies in their design. Researchers did not ask casual users to photograph whatever they happened to eat. Instead, they built a set of 102 meals with precisely measured ingredients, photographed each one under standardized conditions, and fed those images into MyFitnessPal, LoseIt!, CalAI, and Appedi. Every app returned calorie estimates that fell about 250 to 345 calories below actual values, with fat underestimated by about 30 grams. A detailed summary from EurekAlert’s science briefing notes that the consistency of the error across four independent platforms points to a shared weakness in how photo-based AI models interpret food images rather than a flaw unique to any single product.
A separate peer-reviewed study published in npj Digital Medicine reinforces the concern from a different angle. Researchers validated SNAQ, another AI-powered image-based dietary assessment app, against doubly labeled water, the gold standard for measuring total energy expenditure in free-living conditions. The study focused on women with obesity in everyday environments and found systematic underestimation at the group level, along with wide limits of agreement between what SNAQ reported and what participants actually consumed. Doubly labeled water works by tracking metabolic markers in the body over days, making it far harder to fool than a food diary or photo log. The fact that SNAQ’s estimates still fell short under this rigorous comparison suggests the underestimation problem is not limited to lab-controlled meal photos. It persists when real people eat real food in their own homes.
Taken together, the two studies attack the question from opposite ends. The NUTRITION 2026 evaluation controlled for every variable except the AI itself, isolating the model’s accuracy on known meals. The npj Digital Medicine study let participants eat freely and measured total intake through biological markers. Both arrived at the same conclusion: these apps consistently report fewer calories than people actually consume.
Unanswered questions about training data and long-term health effects
Neither study has released the raw datasets or individual meal-level results that would let outside researchers dig into exactly which foods or combinations produce the largest errors. The NUTRITION 2026 evaluation reported only aggregated averages, so it is unclear whether certain cuisines, cooking methods, or portion sizes drove the worst-case 345-calorie gap while others were closer to the 250-calorie floor. Without that granularity, users have no way to know which of their own meals are most likely to be miscounted.
None of the four app developers, MyFitnessPal, LoseIt!, CalAI, or Appedi, have publicly responded to the findings or disclosed details about the image datasets used to train their recognition models. That leaves open basic questions: Are the models skewed toward Western restaurant food? Do they include enough examples of home-cooked stews, mixed plates, or culturally diverse dishes? How often are the databases updated to reflect reformulated products or new fast-food items? Without transparency, it is difficult for clinicians or users to judge whether incremental improvements are happening behind the scenes or whether the same blind spots will persist for years.
The long-term health implications of systematic undercounting are also unknown. For an individual trying to lose weight, a hidden surplus of several hundred calories a day could stall progress, prompting unnecessary changes to medication, exercise routines, or diet plans. For someone with diabetes or cardiovascular disease, chronic underestimation of fat and energy intake could undermine medical nutrition therapy, even when the person believes they are adhering closely to professional advice. At a population level, widespread use of optimistic calorie counts could distort research that relies on app-based food logs, making it harder to link diet patterns to health outcomes.
How to use photo-based food apps without being misled
Despite these limitations, image-based food trackers are unlikely to disappear. They are convenient, engaging, and for many people easier to stick with than traditional food diaries. The challenge is learning how to use them as rough guides rather than precise instruments.
Nutrition experts who work with digital tools often recommend a hybrid strategy. For simple, single-item foods such as a piece of fruit, a plain yogurt, or a grilled chicken breast, the photo-based estimates may be reasonably close. For mixed dishes, creamy sauces, fried foods, and desserts, a more cautious approach helps: manually adjust portion sizes upward, log ingredients separately when possible, or cross-check a few representative meals against a kitchen scale and nutrition labels to get a sense of your personal “error factor.” If you notice that your weight or blood glucose trends do not match what the app suggests should be happening, treat that as a signal to reassess rather than assuming the numbers are correct.
For clinicians and researchers, the emerging evidence argues for clear communication. Patients should be told that these apps can support awareness and pattern recognition-highlighting how often someone eats vegetables or how frequently they snack at night-but that the absolute calorie and fat numbers may be biased low, especially for rich or complex meals. Study designs that rely on app-based intake data may need to incorporate calibration periods or complementary methods, such as occasional weighed food records, to correct for systematic underestimation.
The next generation of digital nutrition tools
The current shortcomings do not mean AI has no role in nutrition; they highlight where the technology needs to evolve. Future models could combine image recognition with contextual cues from barcodes, restaurant menus, and user input about recipes and cooking methods. More diverse and transparently documented training datasets, including home-cooked and culturally varied meals, may reduce bias. Integrating feedback loops-where users can correct obvious errors and those corrections feed back into model updates-could gradually improve accuracy in the wild.
Until then, the message from both controlled experiments and free-living studies is consistent. Photo-based calorie counts are best treated as estimates with a built-in margin of error, not as definitive measurements. For people whose health depends on tight control of energy or fat intake, that margin may be too wide to rely on these apps alone. Used with eyes open, they can still be helpful tools-but they are no substitute for critical thinking, basic nutrition knowledge, and, when needed, professional guidance.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
