Artificial intelligence-powered calorie tracking applications that rely on meal photography are significantly underestimating the energy and fat content of consumed food, according to a new study by researchers at the National Institutes of Health. The findings, which will be presented at the NUTRITION 2026 annual meeting in National Harbor, Maryland, highlight a critical accuracy gap in popular health technology as users increasingly depend on automated dietary logging for weight management. The investigation, led by postdoctoral fellow Aaron Hengist and postbaccalaureate fellow Olivia Charles at the NIH Clinical Center, evaluated four widely used platforms: MyFitnessPal, LoseIt!, CalAI, and Appediet. Unlike previous assessments that relied on casual user-submitted images, the researchers leveraged a controlled metabolic kitchen where every meal component was weighed to the nearest 0.1 gram. Standardized photographs of 102 precisely measured meals were processed through the apps to establish a direct comparison against a verifiable nutritional baseline. The analysis revealed a consistent downward bias across all tested applications. Each platform underestimated total caloric intake by an average of 250 to 345 calories per meal and fat content by approximately 30 grams, representing an underestimation of roughly one-third. While the applications demonstrated improved accuracy for higher-calorie dishes and provided more reliable carbohydrate estimates, they frequently miscalculated other macronutrients. Further examination of over 200 additional meals indicated that algorithmic performance degrades notably with low-carbohydrate, high-fat dietary patterns, likely because the dense fat content of ketogenic meals is systematically undervalued by image recognition models. Researchers emphasize that the convenience of photographic logging comes at the cost of precision. Without manual portion adjustments or supplementary input methods, consumers relying solely on these tools are likely tracking an inflated sense of dietary control. Hengist advised that users should interpret automated photo results with considerable caution, particularly when managing conditions or diet plans heavily reliant on fat intake. To bridge the accuracy gap, the study recommends integrating computer vision features with traditional nutritional logging practices, allowing AI to serve as a supplementary reference rather than a definitive measurement tool. As AI-driven health monitoring continues to expand, the NIH study underscores the necessity for rigorous, standardized validation of consumer-facing algorithms. The researchers will share the complete methodology and performance metrics at the American Society for Nutrition conference later this month, providing a benchmark for developers seeking to improve the clinical reliability of automated dietary assessment technologies.