Nutrients, Vol. 17, Pages 3044: Comparison of Accuracy in the Evaluation of Nutritional Labels on Commercial Ready-to-Eat Meal Boxes Between Professional Nutritionists and Chatbots
Nutrients doi: 10.3390/nu17193044
Authors:
Chin-Feng Hsuan
Yau-Jiunn Lee
Hui-Chun Hsu
Chung-Mei Ouyang
Wen-Chin Yeh
Wei-Hua Tang
Background/Objectives: As convenience store meals become a major dietary source for modern society, the reliability of their nutrition labels is increasingly scrutinized. With advances in artificial intelligence (AI), large language models (LLMs) have been explored for automated nutrition estimation. Aim: To evaluate the accuracy and clinical applicability of AI-assessed nutrition data by comparing outputs from five AI models with professional dietitian estimations and labeled nutrition facts. Methods: Eight ready-to-eat convenience store meals were analyzed. Four experienced dietitians independently estimated the meals’ calories, macronutrients, and sodium content based on measured food weights. Five AI chatbots were queried multiple times with identical input prompts to assess intra- and inter-assay variability. All results were compared to the official nutrition labels to quantify discrepancies and cross-model consistency. Results: Dietitian estimations showed strong internal consistency (CV < 15%), except for fat, saturated fat and sodium (CVs up to 33.3 ± 37.6%, 24.5 ± 11.7%, and 40.2 ± 30.3%, respectively). Among AI models, ChatGPT4.o showed relatively consistent calory, protein, fat, saturated fat and carbohydrate estimates (CV < 15%), and Claude3.7, Grok3, Gemini, and Copilot showed caloric and protein content as consistent (CV < 15%). Sodium values were consistently underestimated across all AI models, with CVs ranging from 20% to 70%. The accuracy of nutritional fact estimation over the five AI models for calories, protein, fat, saturated fat and carbohydrates was between 70 and 90%; when compared to the nutritional labels of RTE, the sodium content and saturated fat estimated were severely underestimated. Conclusions: Current AI chat models provide rapid estimates for basic nutrients and can aid public education or preliminary assessment; GPT-4 outperforms peers in calorie and potassium-related estimations but remains suboptimal in micronutrient prediction. Professional dietitian oversight remains essential for safe and personalized dietary planning.
Background/Objectives: As convenience store meals become a major dietary source for modern society, the reliability of their nutrition labels is increasingly scrutinized. With advances in artificial intelligence (AI), large language models (LLMs) have been explored for automated nutrition estimation. Aim: To evaluate the accuracy and clinical applicability of AI-assessed nutrition data by comparing outputs from five AI models with professional dietitian estimations and labeled nutrition facts. Methods: Eight ready-to-eat convenience store meals were analyzed. Four experienced dietitians independently estimated the meals’ calories, macronutrients, and sodium content based on measured food weights. Five AI chatbots were queried multiple times with identical input prompts to assess intra- and inter-assay variability. All results were compared to the official nutrition labels to quantify discrepancies and cross-model consistency. Results: Dietitian estimations showed strong internal consistency (CV < 15%), except for fat, saturated fat and sodium (CVs up to 33.3 ± 37.6%, 24.5 ± 11.7%, and 40.2 ± 30.3%, respectively). Among AI models, ChatGPT4.o showed relatively consistent calory, protein, fat, saturated fat and carbohydrate estimates (CV < 15%), and Claude3.7, Grok3, Gemini, and Copilot showed caloric and protein content as consistent (CV < 15%). Sodium values were consistently underestimated across all AI models, with CVs ranging from 20% to 70%. The accuracy of nutritional fact estimation over the five AI models for calories, protein, fat, saturated fat and carbohydrates was between 70 and 90%; when compared to the nutritional labels of RTE, the sodium content and saturated fat estimated were severely underestimated. Conclusions: Current AI chat models provide rapid estimates for basic nutrients and can aid public education or preliminary assessment; GPT-4 outperforms peers in calorie and potassium-related estimations but remains suboptimal in micronutrient prediction. Professional dietitian oversight remains essential for safe and personalized dietary planning. Read More