Journal of Global Agriculture and Ecology

Artificial intelligence is entering agricultural extension at a moment when the field has already moved away from linear technology transfer towards pluralistic, participatory and innovation-systems approaches. This review critically examines whether artificial intelligence supports that transition or merely automates an older delivery model. It compares low- and middle-income countries with high-income countries, treating these groupings as contrasting configurations of infrastructure, farm structure, data availability, advisory institutions and political economy rather than as a simple development sequence. A transparent narrative search and appraisal process was used to integrate evidence on digital advisory services, agricultural knowledge and innovation systems, precision agriculture, machine learning, conversational tools, adoption, inclusion, data governance and responsible innovation. The evidence indicates that artificial intelligence can extend diagnostic reach, personalise advice, translate information, prioritise cases and connect farmers with services. Yet the strongest empirical evidence still concerns relatively simple digital information services, while claims about autonomous or generative systems often exceed evidence on sustained farm-level outcomes. In low- and middle-income countries, mobile-first and human-mediated models can broaden access but remain constrained by connectivity, language, literacy, gendered access, weak local datasets and project dependence. In high-income countries, sensor-rich systems offer greater technical integration but intensify concerns about interoperability, capital bias, proprietary platforms, data rights and farmer dependence. Across both contexts, performance depends less on model novelty than on agronomic validation, institutional fit, uncertainty communication, human escalation and accountable data governance. The review proposes a human–artificial intelligence extension ecology in which extension professionals act as contextual interpreters, validators, facilitators and data stewards rather than passive transmitters or displaced intermediaries. Artificial intelligence is therefore most defensible as an augmentative public-interest infrastructure embedded in pluralistic advisory systems, not as a substitute for relational extension or farmer judgement.