Artificial Intelligence in Animal Husbandry: A Critical Narrative Review of Diagnostic Performance, Methodological Quality and Translational Constraints

Abstract

Artificial intelligence has become central to the technological narrative surrounding modern animal husbandry, yet the strength of the underlying evidence is uneven and frequently overstated. This critical narrative review examines what can defensibly be concluded about machine learning and deep learning applications across cattle, pig, poultry, sheep and goat production systems, and where confidence remains weak. Literature was identified through six scholarly sources supplemented by citation searching and examination of authoritative institutional publications, with the final search conducted on 23 June 2026. Evidence was appraised for design adequacy, reference-standard quality, validation strategy, metric selection, external validity and reporting transparency, and was synthesised thematically rather than study by study. Three findings dominate the synthesis. Performance figures reported for algorithmic detection of health and reproductive events are systematically optimistic because validation designs frequently permit dependency between training and evaluation data, because reference standards for outcomes such as lameness and subclinical mastitis are themselves imperfect and inconsistently defined, and because discrimination metrics are reported in place of decision-relevant measures such as predictive value under realistic prevalence. Second, the evidence base is architecturally deep but epidemiologically shallow: algorithmic novelty is abundant, whereas independent external validation across farms, breeds, housing systems and climates is scarce, and the small number of studies that have tested transportability report substantial performance loss. Third, the constraints that most limit practical benefit are organisational rather than computational, encompassing data governance, interoperability, security, workforce competence and the concentration of both technology and research in high-income intensive systems. Claims of demonstrated effectiveness are therefore warranted for a narrow set of well-characterised tasks, principally activity-based oestrus detection and automated body measurement, and are not warranted for most welfare inference. Priorities include standardised case definitions, prospective multi-farm validation, economically anchored evaluation, and deliberate attention to smallholder and tropical production systems.