Agricultural Horizons: Research, Technology and Sustainable Development Vol. 4

Human-AI Collaboration in Agricultural Extension: Reconfiguring Extension-Agent Roles in the Generative AI Era

L. Raja — 2026-09-29

Generative artificial intelligence (GenAI), especially large language models (LLMs), is entering agricultural extension at a point when advisory systems are already being reshaped by digitalisation, pluralistic service provision and data-intensive agriculture. This critical narrative review examines how human-AI collaboration changes the work, authority and accountability of extension agents rather than treating GenAI as a direct substitute for human advisory labour. Literature published from 1 January 2000 to 4 July 2026 was considered, with earlier foundational studies retained where necessary to interpret human-automation relationships. Evidence was integrated across agricultural extension, digital agriculture, human-AI collaboration, trust, data governance and emerging empirical evaluations of LLM-enabled agricultural advice. The strongest current evidence supports GenAI as an augmentation layer for information retrieval, drafting, translation, question triage and first-line decision support. Field and expert-evaluation studies also show that model outputs can be technically useful, yet remain vulnerable to local agronomic mismatch, incomplete contextualisation, variable expert ratings and accountability concerns. Evidence that GenAI independently improves long-term farm productivity, income, resilience or equity remains limited. The review therefore argues for a shift from an information-delivery conception of extension work to a bounded-delegation model in which extension agents act as evidence curators, agroecological contextualisers, trust calibrators, facilitators of social learning, data stewards and accountable escalation points. Organisational design is as important as model capability: retrieval-grounded knowledge bases, risk-tiered task allocation, auditability, feedback loops, professional training and public-interest data governance are required to prevent automation bias, deskilling and exclusion. Human-AI collaboration can strengthen extension where it redistributes routine cognitive work towards machines while protecting the relational, interpretive and institutional functions that make advice locally actionable and socially legitimate.

Biofloc Aquaculture 2.0: A Critical Appraisal of Microbial Community Engineering, Resource Efficiency and Disease Resilience

Hariom Bohare — 2026-09-29

Biofloc technology has moved within two decades from an empirical pond-management practice to a component of intensive, low-exchange production for penaeid shrimp and warm-water finfish. The original control logic rested on a single manipulated variable, the carbon-to-nitrogen ratio of material entering the culture unit, and on the assumption that heterotrophic assimilation of ammonium would dominate nitrogen transformation. High-throughput sequencing, genome-resolved metagenomics and quantitative synthesis have since produced a more complicated picture, and a second-generation framing has emerged in which the microbial community itself is treated as the object of engineering rather than as an incidental consequence of carbon dosing. This review critically evaluates whether that framing is supported by current evidence. Literature was identified through structured searching of seven scholarly databases and indexes, supplemented by citation tracking and an intergovernmental source, covering 1999 to 14 July 2026. Four propositions were examined: that biofloc communities have a reproducible core amenable to directed manipulation; that carbon management delivers predictable gains in nitrogen and feed resource efficiency; that disease resilience arises from identifiable and controllable mechanisms; and that system-level resource efficiency extends beyond water saving. The evidence is strongest for a taxonomically recurrent, functionally versatile floc community, for partial and size-dependent nutritional contribution of flocs, and for pathogen-directed virulence attenuation under defined conditions. It is weakest for predictive control of community assembly, for long-term nitrogen closure given documented nitrate accumulation, and for energy and greenhouse-gas performance, which remain largely unquantified for biofloc units specifically. Reports of human-associated pathogens and intrinsic antimicrobial resistance in poorly managed freshwater systems indicate that resilience is conditional rather than inherent. Progress requires standardised reporting, genome-resolved rather than amplicon-only characterisation, longer trials at commercial scale, and explicit falsification tests of the engineering claim.

Precision and Digital Farming Reconsidered: A Critical Narrative Review of Sensors, Artificial Intelligence and Remote Sensing in Crop Production

Syed Shujat Hussain — 2026-09-29

Sensing hardware, satellite and unmanned aerial observation, and machine learning are now routinely presented as the technological foundation of a more productive and less environmentally damaging agriculture. The scholarly literature supporting this position has expanded rapidly, yet it remains fragmented across agronomy, remote sensing, computer science, agricultural economics and rural sociology, and the strength of the underlying evidence is uneven. This review examines the state of knowledge on sensing, artificial intelligence and remote sensing in arable and horticultural crop production, and evaluates how far demonstrated technical performance has translated into demonstrated agronomic, economic and environmental benefit. Literature was identified through structured searching of open bibliographic indexes and scholarly search platforms, supplemented by citation searching of recent reviews and by targeted retrieval of institutional reports, with every retained source verified against the digital object identifier registry. The synthesis is organised around four analytical layers: measurement, interpretation, decision and actuation. The evidence is strongest for measurement, where sensor and platform performance has been characterised extensively, and weakest at the point where measurements become agronomic decisions. Machine learning models frequently achieve high accuracy under conditions closely resembling their training data, but independent evaluation across sites, seasons and cultivars is uncommon, and reported accuracies are therefore poor predictors of operational value. Field evidence for variable-rate nutrient management and for autonomous weeding is more heterogeneous than technology-focused reviews imply, and environmental claims frequently rest on short-duration trials concentrated in industrialised temperate systems. Persistent constraints on adoption are institutional and distributional rather than technical, involving data governance, trust, interoperability and unequal access to connectivity and ground data. The principal unresolved questions concern the transferability of models, the agronomic response functions on which prescriptions depend, and the conditions under which digital tools benefit smallholder systems.