Generative Artificial Intelligence in Agricultural Extension: A Critical Review of Applications, Evidence Limitations and Research Priorities

Abstract

Agricultural extension and advisory systems face a long-standing mismatch between the number of farmers requiring individualised technical guidance and the capacity of public and private advisory organisations to supply it. Generative artificial intelligence, and large language models in particular, has been proposed as a means of narrowing that gap by producing conversational, multilingual and multimodal advice at low marginal cost. Enthusiasm has outpaced evidence. This critical review examines what is currently known about generative artificial intelligence in agricultural extension, separating demonstrated capability from inferred benefit. Literature was identified through structured searching of open scholarly indexes and metadata registries, supplemented by backward and forward citation tracking and by searches of institutional repositories, and was appraised for methodological adequacy rather than accepted on grounds of topical relevance alone. Four application clusters emerged: conversational advisory services aimed directly at farmers; multimodal diagnostic support for pests and diseases; augmentation of extension professionals and content production; and automated construction of agricultural knowledge resources. Evidence across these clusters is markedly asymmetric. Performance on retrieval and question-answering tasks is documented reasonably well, evidence on agronomic accuracy under field conditions is thin, and evidence on farmer-level outcomes such as practice change, yield or income is close to absent. Retrieval augmentation reduces but does not remove factual error, and prevailing benchmarks reward fluency and source fidelity rather than agronomic appropriateness. Studies that disaggregate performance by question type indicate systematic weakness precisely where advice must be site specific, notably input rates and timing. Constraints relating to language coverage, digital literacy, connectivity, gendered access and agroecological localisation are widely acknowledged yet rarely measured. Governance questions concerning liability for erroneous advice, data rights and advisory dependence remain unresolved. Research priorities include agronomically grounded evaluation protocols co-developed with domain specialists, experimental assessment of advisory outcomes, systematic measurement of performance in low-resource languages, and empirical study of how advisers and farmers actually incorporate machine-generated recommendations into decisions.