Chemistry-informed AI for food toxicant and adulterant detection: matrix-aware validation and auditable decision support

Food toxicants and adulterants pose interconnected safety, authenticity, and regulatory challenges across complex food matrices. This review examines how artificial intelligence (AI) converts spectra, images, chromatographic-mass-spectral fingerprints, biosensor outputs, and portable-device signals into screening decisions. Toxicant detection is treated primarily as a threshold-based task, whereas adulteration is addressed as an authenticity or anomaly-detection problem. Evidence is evaluated using AI performance and classical analytical criteria, including selectivity, recovery, precision, robustness, limits of detection and quantification, measurement uncertainty, matrix effects, independent validation, and agreement with reference methods. The review distinguishes direct analyte evidence from correlated proxy signals and separates experimentally demonstrated methods from author-proposed chemistry-informed concepts. A matrix-aware framework links field screening with confirmatory laboratory analysis through calibration transfer, drift monitoring, uncertainty-based decisions, and traceable model records. Emerging tools, including conformal prediction, generative augmentation, digital twins, and secure ledgers, remain promising but are not yet fully validated for regulatory use in foods.