Knowledge, Innovation and Technology in Scientific Research Vol. 6

Artificial intelligence (AI), particularly generative AI and large language models, is entering agricultural extension at a point when advisory systems are already under pressure to provide timely, locally specific knowledge at scale. This review critically examines responsible AI in agricultural extension through five connected domains: accuracy, accountability, bias, privacy and farmer protection. A critical narrative approach was used to integrate evidence from agricultural extension, digital agriculture, AI ethics, data governance and recent evaluations of AI-enabled advisory tools. The literature indicates that AI can widen access to information, assist extension personnel and support localisation of technical content, but the evidence base remains stronger for feasibility and expert-rated response quality than for sustained farmer-level outcomes. Accuracy is context dependent: fluent answers can remain agronomically unsafe when planting dates, input rates, disease-control options or local regulations are wrong or outdated. Accountability is frequently diffuse across foundation-model providers, application developers, knowledge curators, extension organisations and human advisers. Bias can arise from unequal connectivity, under-represented languages and agroecologies, gendered assumptions, farm-size effects and commercial incentives, while aggregate accuracy can conceal unequal harms. Privacy concerns extend beyond conventional personal data to economically sensitive farm, location, production and behavioural information whose value may emerge through inference or linkage. The review argues that farmer protection should be treated as an operational safety-and-governance obligation rather than as a disclosure exercise. Responsible deployment therefore requires locally grounded evidence, risk-tiered human oversight, subgroup testing, data minimisation, traceability, incident management, accessible complaints and meaningful remedies. A proposed extension-system governance model links evidence governance, model governance, deployment safeguards, institutional accountability and farmer-facing protections. The principal research need is independent, longitudinal and multi-context evaluation that measures not only technical performance but also safety incidents, distributional effects, trust, autonomy, economic outcomes and the quality of human-AI extension relationships.