Abstract
Small and marginal farmers face intersecting challenges related to food security, environmental risk and structural disadvantage. Agricultural extension has historically played a central role in supporting these farmers, with evolving approaches that increasingly emphasise participatory learning, farmer agency and ethical knowledge exchange. As artificial intelligence (AI) technologies begin to enter the agricultural advisory landscape, their potential to support smallholder learning remains both promising and contested. This article explores the intersection of AI and agricultural extension by proposing a typology of learning based on two key dimensions: the locus of knowledge production and the orientation of agricultural knowledge and innovation systems (AKIS). 
Using this framework, we assess the extent to which current AI applications in agriculture align with ethical and participatory extension goals. Our analysis is grounded in a detailed case study of Farmer.Chat, a generative AI-powered advisory tool developed by Digital Green and Microsoft Research, and deployed in four countries. Drawing on mixed-methods data, we examine how AI can support or limit different types of learning, trust-building and knowledge co-creation. We find that while Farmer.Chat enhances access and personalisation, it still leans towards individualised, one-way communication. Its full potential depends on embedding it within trusted social infrastructures, enabling feedback loops and aligning with double-loop learning and participatory extension ethics. We conclude with a research agenda to guide the development of AI tools that support more inclusive, adaptive and democratic agricultural knowledge systems.
Using this framework, we assess the extent to which current AI applications in agriculture align with ethical and participatory extension goals. Our analysis is grounded in a detailed case study of Farmer.Chat, a generative AI-powered advisory tool developed by Digital Green and Microsoft Research, and deployed in four countries. Drawing on mixed-methods data, we examine how AI can support or limit different types of learning, trust-building and knowledge co-creation. We find that while Farmer.Chat enhances access and personalisation, it still leans towards individualised, one-way communication. Its full potential depends on embedding it within trusted social infrastructures, enabling feedback loops and aligning with double-loop learning and participatory extension ethics. We conclude with a research agenda to guide the development of AI tools that support more inclusive, adaptive and democratic agricultural knowledge systems.
