Abstract

TITLE: When Food Waste Quantification Falls Short: An Exploratory Study of AI-Assisted Measurement

AUTHORS: Mengting Yu1, Luca Secondi1, Luigi Palumbo1, Clara Ciatiello1, Tiziana Laureti11, Ludovica Principato2, Camila Comis1
1. University of Tuscia, Viterbo (Italy)
2. Roma Tre University, Rome (Italy)

Abstract:
Food waste (FW) remains a critical yet insufficiently explored global challenge, and accurate quantification is a prerequisite for designing, evaluating, and scaling effective mitigation strategies. Existing FW measurement approaches, such as self-reported surveys, manual waste tracking, and compositional analyses, have generated valuable insights. Still, they suffer from well-documented limitations related to accuracy, scalability, cost-efficiency, and respondent burden. These constraints can introduce systematic biases and misjudgments in FW estimates, particularly in free-living consumption contexts, thereby weakening the empirical basis for policy evaluation and intervention design.

Recent advances in artificial intelligence (AI), especially in image analysis, offer new opportunities to address these challenges. AI technologies enable automated, continuous, and standardized data collection, reducing reliance on subjective reporting and facilitating large- scale monitoring across households, food services, and retail environments. Within food science, AI applications, most notably Large Language Models (LLMs) combined with vision capabilities, have demonstrated promising outcomes in areas such as food recognition, nutritional assessment, quality inspection, and dietary planning. Despite these successes, empirical applications of AI to FW quantification remain scarce, and there is limited understanding of their reliability, error structures, and practical constraints in this domain.

Addressing this gap, this exploratory study examines the potential and limitations of AI-driven FW quantification using LLM-based image analysis. The methodological approach compares AI-estimated and human-estimated FW weights against objectively measured weights obtained using kitchen scales, allowing for a direct benchmark of accuracy and error patterns. Food images were collected in unconstrained, real-world settings, reflecting diverse cuisines, containers, angles, depth, and lighting conditions. Multiple open-source LLMs with varying parameter sizes were employed to evaluate model performance and robustness.

The results indicate that AI-assisted FW quantification holds considerable promise as a scalable and automated measurement tool, but performance varies substantially across models. This variability highlights the importance of careful model selection and suggests that further training with larger and more diverse datasets could substantially enhance accuracy. At the same time, the findings underscore current limitations of LLM-based estimation, particularly in relation to volume and weight inference from two-dimensional images.

Overall, this study represents an early contribution to the measurement and monitoring of FW using AI technologies and, to our knowledge, the first systematic applications of LLMs in FW identification and quantification. The results point to significant potential for AI-enabled data streams to complement traditional methods, while emphasizing the need for continued methodological validation and model development to support robust, large-scale FW quantification.