Artificial intelligence in dietary assessment: a narrative review

Abstract Objectives To evaluate the accuracy, applicability, and limitations of artificial intelligence-based dietary assessment (AI-DA) tools developed between 2015 and 2025. Background AI-DA has emerged as a promising solution to the limitations of traditional self-report dietary methods, which are prone to recall bias, under-reporting, and high respondent burden. Advances in computer vision, Natural Language Processing (NLP), wearable sensors, and multimodal AI offer new opportunities for automated dietary assessment. Methods A structured literature search was conducted using PubMed, Scopus, Web of Science, and Google Scholar to identify validation studies, feasibility studies, observational research, and relevant systematic and scoping reviews published between 2015 and 2025. Evidence was synthesized narratively, focusing on quantitative accuracy, feasibility in free-living settings, and methodological limitations. Results Computer-vision models achieve high accuracy for food classification (approximately 70%-95%). NLP-enabled automated recalls decrease interviewer burden and improve reporting consistency. Wearable and sensor-based systems enable passive monitoring of eating behaviours, while multimodal AI systems integrating images, text, and sensor data enhance nutrient estimation. Persistent challenges include unreliable portion-size estimation especially for mixed or amorphous foods—culturally narrow training datasets, environmental variability, and algorithmic bias. Additional barriers include privacy concerns, restricted interoperability with digital health systems, and dependence on reliable internet connectivity. Conclusions AI-DA tools show strong potential for enhancing dietary assessment in research, clinical care, and public health. Future progress requires culturally diverse datasets, standardized validation frameworks, and improved integration with digital health systems.