Built for Farmers, Tested Without Them: The Blind Spot in Agricultural AI

Artificial intelligence (AI) is moving quickly into agriculture, promising earlier disease detection, more accurate yield forecasts, better irrigation decisions and faster access to farm advice. However, its progress hides a fundamental problem: much of the research is being conducted around smallholder farmers rather than with them, says a new study published in the MDPI journal Agriculture.

Titled "Technology-Driven or Farmer-Centred? A Systematic Review of Artificial Intelligence Research for Smallholder Agriculture," the study examines whether the rapid expansion of agricultural AI is translating into tools that small-scale farmers can actually use. The authors reviewed 182 studies and found a field rich in technical experimentation but thin in evidence about adoption, affordability and practical value. While 178 studies applied AI to small-scale farming problems, only nine examined how farmers used, perceived or engaged with the technology. Five appeared in both categories, reinforcing the scale of the imbalance rather than narrowing it.

AI boom is being shaped by what machines can easily measure

The study captures a research field that has expanded dramatically in a short period. Publication activity was limited before 2020, accelerated from 2021 and reached its highest level in 2025. India accounted for 30 of the reviewed studies, China for 25 and South Africa for 20, with Ethiopia, Tanzania, Kenya, Nigeria and Bangladesh also emerging as important research locations.

This geographic pattern reflects a mix of agricultural need and research capacity. Africa and Asia contain large smallholder populations and face acute food-security and climate pressures, but the strongest concentrations of research appear in countries with comparatively developed universities, digital-agriculture programmes, data systems and technology sectors.

Smallholder agriculture is highly diverse. A model developed for maize production in one agroecological zone may perform poorly where soils, languages, cropping practices, infrastructure or market conditions differ. Yet many countries with substantial smallholder populations remain underrepresented in the evidence base.

The thematic imbalance is equally revealing. Plant-disease detection, crop-yield prediction, crop classification, land mapping and environmental monitoring dominate the literature. Convolutional neural networks are widely used for image-based applications, while Random Forest and other tree-based methods are common where structured agricultural data are available.

Earlier disease diagnosis can reduce losses, while improved yield estimates can support food planning, insurance, credit and input allocation. But their dominance also suggests that research agendas are being shaped by data availability and technical convenience.

Plant leaves can be photographed. Fields can be mapped by satellite. Crop yields can be modelled using weather, soil and vegetation data. By comparison, farmer trust, household risk, local knowledge, gender relations, market power and willingness to adopt are harder to quantify. As a result, the research tends to favour problems that fit existing AI methods rather than the full range of problems that shape smallholder livelihoods.

Areas such as livestock health, post-harvest losses, market access, integrated soil and water management, farm finance and whole-farm decision-making receive far less attention. The danger is that AI becomes highly sophisticated at solving a narrow set of measurable problems while overlooking the economic and institutional constraints that determine whether better information changes outcomes.

The missing variable in agricultural AI is the farmer

Only nine studies explored adoption, perceptions, behavioural responses or practical use, compared with 178 that applied AI as a research, analytical or decision-support tool. Even within that small group, the intensity of engagement varied sharply. Some farmers actively used or piloted hydroponic systems, cassava-disease applications and the MaizeNet mobile tool. Others participated as beta testers, evaluators or design consultants. Several studies examined attitudes through interviews or hypothetical scenarios without deploying an AI tool in everyday farming.

Technical accuracy is only one part of a functioning agricultural service. A disease-detection application may achieve impressive laboratory results but still fail if farmers lack compatible smartphones, cannot afford data, do not recognize the language used, distrust the recommendation or cannot purchase the suggested treatment. A yield-prediction system may improve forecasting without improving farmers' bargaining power, income or access to markets.

The review exposes a larger weakness in the digital-development debate: intended beneficiaries are often treated as end users rather than co-designers. Their role begins after the problem has been selected, the model built and the interface designed.

A genuinely farmer-centred approach would start earlier. It would ask farmers which decisions generate the greatest uncertainty, how they currently obtain information, what forms of advice they trust and what costs they can realistically bear. It would also examine how technology affects different groups, including women farmers, tenant farmers, pastoralists, older farmers and communities with limited literacy or connectivity.

For governments and development institutions, this means success should not be measured by the number of pilots launched or models trained. More meaningful indicators would include continued use, affordability, accuracy under local conditions, reduced losses, improved income, stronger resilience and the extent to which farmers can challenge or verify an AI-generated recommendation.

The review's framing is particularly important for the Global South. Smallholders often operate in environments where extension services are understaffed, data are fragmented and connectivity is unreliable. AI may help close information gaps, but it could also deepen inequality if the best tools remain available only to farmers with better devices, stronger networks, formal land records or access to commercial platforms.

The real challenge is moving from clever prototypes to durable farm systems

The study finds that most agricultural AI remains closer to model development than sustained implementation. Controlled testing, technical validation and simulated environments are common, while long-term evidence on field use, maintenance, farmer feedback and livelihood outcomes is scarce.

The gap between prototype and practice is where many development technologies falter. An AI tool must function not only when researchers are present but also after a pilot ends. It must survive poor connectivity, damaged devices, changing crop conditions, incomplete data and limited technical support. It must remain useful when farmers face multiple, simultaneous decisions involving production, weather, labour, credit, prices and household needs.

The authors therefore call for a shift from isolated applications towards integrated decision-support systems that combine production, climate, financial and market information. They also emphasize collaboration among researchers, technology developers, extension services, policymakers and farmers, alongside stronger infrastructure, digital literacy and participatory design.

This has practical implications for investment. Governments and donors should make field validation and farmer participation conditions of public support, rather than optional additions to technically promising projects. Procurement decisions should consider offline functionality, language access, data protection, interoperability and long-term maintenance.

Businesses also have an opportunity, particularly in developing lightweight tools that can run on lower-cost devices and work with intermittent connectivity. But commercial scale should not be mistaken for social value. A profitable platform may still exclude poorer farmers, lock users into proprietary ecosystems or extract agricultural data without providing adequate control or compensation.

Extension services will remain crucial. AI is unlikely to replace trusted intermediaries in the near term, particularly where recommendations involve financial risk or complex local conditions. A more realistic model may combine machine-generated analysis with human interpretation, allowing extension workers to validate outputs and adapt advice to local circumstances.

Agricultural AI needs a new definition of progress

The review followed PRISMA 2020 guidelines, searched Web of Science, Scopus and EBSCOhost, screened an initial 8,600 records and retained 182 studies after duplicate removal and eligibility assessment. The authors also conducted a formal quality appraisal of all included studies.

Its limitations, however, should temper broad conclusions. The search was restricted to English-language publications indexed in three databases. Relevant regional studies, non-English research, theses, book chapters and inaccessible full texts may therefore have been missed. The fast pace of AI development also means newer approaches may not be fully represented, while classifying studies as researcher-driven or farmer-centred inevitably involved judgement.

Despite the limitations, the study delivers a powerful strategic message. Agricultural AI cannot be judged only by whether it predicts more accurately. It must also be assessed by who shapes it, who controls the data, who can access the resulting service and who bears the cost when advice is wrong.

For policymakers, the next stage of agricultural digitalization should therefore focus less on showcasing innovation and more on building the conditions for responsible use. It includes investment in rural connectivity, local-language interfaces, farmer training, extension capacity, public-interest data governance and independent evaluation.

Researchers must test technologies over longer periods and measuring outcomes that matter beyond the laboratory: adoption, trust, income, resilience, labour burdens and distributional effects. It also means investigating whether simpler, cheaper and more explainable models may sometimes be more useful than technically advanced systems.