One Health, One Genome: A Critical Appraisal of Artificial Intelligence–enabled Genomics for Global Infectious Disease Surveillance

Pathogen genomics has become a routine instrument of infectious disease surveillance, and machine learning is increasingly proposed as the means by which sequence data generated across human, animal and environmental sectors can be converted into anticipatory public health intelligence. The premise that a single integrated genomic evidence base can serve all three sectors, and that artificial intelligence can extract predictive signal from it, has attracted substantial investment, yet the supporting evidence remains uneven and has not been appraised critically as a whole. This review evaluates the strength, consistency and limitations of the literature on artificial intelligence–enabled genomics for One Health infectious disease surveillance, covering lineage assignment and phylogenetic automation, sequence-based fitness and antigenic escape prediction, protein and genome language models, cross-species host and spillover inference, genotype-to-phenotype prediction of antimicrobial resistance, and wastewater and environmental metagenomics. Literature was identified through Europe PMC, Crossref and the Directory of Open Access Journals, supplemented by institutional sources, with all bibliographic records and digital object identifiers verified against registration metadata. The evidence is strongest where algorithms perform structured classification against well-curated reference data, notably clade and lineage assignment and resistance determinant detection in taxa with dense phenotype-linked genome collections. Confidence weakens progressively for retrospective fitness inference, and is weakest for prospective cross-species risk prediction, where reported discrimination is difficult to separate from sampling bias in the underlying host–virus association records. Independent reanalyses indicate that apparent predictive skill often reflects research effort and taxonomic structure rather than transferable biological signal. Recurrent methodological problems include data leakage across phylogenetically related sequences, absent external validation, severe geographical concentration of both genomes and metadata, and an almost complete lack of evaluation against public health decision outcomes rather than classification metrics. Progress depends less on model architecture than on representative sampling across sectors, interoperable contextual metadata, prospective evaluation designs, and governance arrangements that address equity and dual-use risk simultaneously.