Smallholder farmers in developing countries generate rich, structured agricultural knowledge through social media platforms every day — crop diagnostic descriptions, peer-validated price information, technique queries, and market intelligence — without knowing they are doing so, without owning what they produce, and without governance frameworks that protect their rights over this data or ensure they benefit from its potential use in artificial intelligence systems. This paper examines farmer-generated social media data from an original survey of 720 smallholder farmers across all 20 administrative blocks of Prayagraj district, Uttar Pradesh, India, and applies the Responsible AI and Data Justice frameworks to evaluate the ethical conditions under which AI agricultural advisory systems could legitimately build on this data. We find that the data farmers generate constitutes a structurally valuable resource for training localised AI advisory systems — including labelled diagnostic image-text pairs, real-time hyperlocal price signals, and agricultural information demand data — but that current governance conditions are systematically inadequate: genuine informed consent is absent, community data ownership does not exist, language representation is ignored, and algorithmic accountability mechanisms are non-existent. We argue that the choice for agricultural AI development in the Global South is not between using farmer-generated data or not, but between using it extractively and using it responsibly. Realising responsible AI agricultural advisory requires not technical solutions but governance ones: community data trusts, participatory consent mechanisms, language justice requirements, algorithmic bias auditing, and public institutional architecture that aligns AI development with farming community interests rather than technology corporate extraction.