Artificial Intelligence and Large Language Models in Agricultural Extension: Critical Synthesis of Adoption, Efficacy, and Equity in Smallholder Advisory Systems

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Agricultural extension and advisory services are undergoing a rapid digital transformation driven by artificial intelligence (AI), machine learning, natural language processing, and, more recently, generative artificial intelligence and large language models (LLMs). These technologies have the potential to address persistent constraints in conventional agricultural extension, including limited adviser-to-farmer ratios, geographical isolation, inadequate localisation of information, and delays in the dissemination of climate, market, and agronomic information. LLM-based advisory systems are particularly significant because they can transform complex agricultural knowledge into conversational, multilingual, and potentially personalised guidance. However, their technical capability does not automatically translate into effective or equitable agricultural outcomes. This review synthesises emerging evidence on the adoption, efficacy, scalability, and equity implications of AI and LLM-based agricultural advisory systems, with particular emphasis on smallholder farmers in low- and middle-income countries. The literature indicates that AI-enabled advisory systems can improve access to timely information, facilitate question-answering, support crop and pest management, and extend the reach of human extension personnel. Emerging platforms such as Farmer.Chat demonstrate the feasibility of large-scale generative AI-enabled advisory delivery, while recent experiments with retrieval-augmented generation indicate promising pathways for improving factual grounding and contextual relevance. Nevertheless, evidence of direct impacts on farm productivity, profitability, resilience, and sustained behavioural change remains considerably weaker than evidence of technological feasibility and user engagement. Major limitations include hallucination, outdated or geographically inappropriate information, language and dialect constraints, weak interoperability with local agricultural databases, limited transparency, data governance concerns, and overreliance on automated recommendations. Equity is an equally important concern because digital access, literacy, gender, income, connectivity, and social norms influence who can benefit from AI-enabled services. The review argues that LLMs should therefore be positioned as augmentation technologies within pluralistic extension systems rather than as substitutes for human advisers. Future agricultural AI systems should combine retrieval from authoritative local knowledge bases, multimodal and multilingual interfaces, human oversight, participatory design, continuous evaluation, responsible data governance, and explicit equity metrics. The future value of AI in agricultural extension will ultimately depend not on model sophistication alone but on the quality of the socio-technical systems in which these technologies are embedded.