Beyond Accuracy: Validation, Reproducibility, and Deployment Readiness of AI in Food Safety and Quality

Artificial intelligence (AI) is increasingly proposed for detecting microbial and chemical hazards in food, but laboratory accuracies are rarely reproduced in routine use. This structured narrative review examined whether reported performance is traceable, independently validated, and transferable to operational settings. Eight bibliographic databases and seven institutional routes were searched for English-language records published from January 2018 to May 2026, and we screened the records; the final corpus comprised 76 sources (53 journal articles, 9 conference papers, 2 preprints, and 12 institutional or regulatory documents). Seven primary studies supporting the quantitative evidence were appraised individually against 11 indicators, including test-set independence, external validation, and leakage control. Physical-hazard inspection and narrow-matrix spectral grading had the strongest deployment evidence, whereas microbial and chemical screening remained at laboratory or pilot scale. Among the appraised studies, the most credible estimates came from physically or temporally independent test sets, and no sensor-based study evaluated transfer to another instrument or laboratory. Deep learning showed clear advantages for image-based tasks but not consistently for small-sample spectral calibration. Adoption is limited mainly by validation design, recalibration burden, false-alarm costs, and the absence of a regulatory pathway rather than by accuracy. AI is best positioned as a screening and decision-support layer linked to confirmatory testing and human oversight.