Artificial Intelligence Applications in Climate-Smart Agriculture: A Critical Review of Evidence, Implementation and Responsible Innovation

Abstract

Climate-smart agriculture seeks to increase agricultural productivity and livelihoods, strengthen adaptation and resilience, and reduce or remove greenhouse-gas emissions where feasible. Artificial intelligence is increasingly presented as a means of operationalising these aims through prediction, diagnosis, optimisation and automation. Yet the literature is fragmented across agronomy, remote sensing, engineering, computer science and rural social science, and technical accuracy is often treated as a proxy for climate-smart impact. This critical narrative review evaluates how artificial intelligence contributes to climate-smart agriculture, where the evidence is strongest, and which methodological and institutional constraints limit reliable translation. Publicly accessible scholarly indexes, repositories and verified DOI records were searched for literature published from 1 January 2006 to 30 May 2026, with selected foundational studies included. Evidence was synthesised across sensing infrastructures, yield and climate-risk prediction, crop and livestock health, irrigation and nutrient management, soil and carbon assessment, robotics, decision support, and governance. The evidence is strongest for well-bounded perception and prediction tasks with abundant labelled data, including image-based disease recognition, crop mapping and some yield-estimation applications. Confidence weakens when models are transferred across seasons, regions, cultivars and management systems, or when claimed benefits depend on unmeasured behavioural, economic or environmental responses. Many studies rely on random data splits, narrow benchmark datasets and retrospective accuracy metrics, while relatively few establish causal effects on water use, emissions, profitability, resilience or distributional outcomes. Hybrid approaches that combine process knowledge, spatially structured validation, uncertainty communication and human oversight are more defensible than unconstrained black-box deployment. Artificial intelligence can therefore support climate-smart agriculture, but it is not inherently climate-smart. Its contribution depends on data representativeness, agronomic validity, energy and material costs, interoperability, farmer agency, accountable governance and evaluation against all three climate-smart objectives. Future progress requires multi-location prospective trials, transparent reporting, locally governed data infrastructures and outcome-based assessment that treats equity and ecological effects as core performance criteria