AI-driven approaches for food risk assessment: current advances and future trends

Artificial intelligence (AI) is increasingly being explored as a tool to support food risk assessment and broader food safety activities in response to the growing complexity of global food supply chains, emerging foodborne hazards, and evolving food fraud practices. Unlike previous reviews that have largely focused on individual AI technologies or specific applications, this review provides a comprehensive and critical synthesis of machine learning (ML), deep learning (DL), transformer-based models, explainable artificial intelligence (XAI), multimodal learning, and emerging AI paradigms within a unified framework relevant to food risk assessment and food safety decision-making. The review examines their applications in hazard identification, contamination detection, predictive microbiology, risk classification, food fraud detection, surveillance, traceability, and decision-support systems. Particular attention is given to the strengths, limitations, interpretability, validation requirements, and implementation challenges associated with different AI approaches. The integration of AI with multimodal data sources, Internet of Things (IoT) technologies, spectroscopy, biosensors, and advanced analytical platforms is also discussed as a means to improve risk prediction, monitoring, and early-warning capabilities. Key challenges include data quality and availability, model generalizability, limited external validation, computational requirements, interpretability, and regulatory acceptance. Emerging developments, including foundation models, federated learning, blockchain, digital twins, edge AI, and multimodal frameworks, are highlighted as promising directions for future research. Overall, this review provides a critical comparative framework for understanding the evolving role of AI in food risk assessment and related food safety applications, while identifying key scientific, technical, and regulatory priorities for future development and implementation. Graphical Abstract