Identifying systemic risks and mitigation strategies of artificial intelligence in agriculture: from social-technical

AI in agriculture, including using AI in planting, cropping, and other agricultural production, faces various forms of risk across its technical, social, and ecological subsystems. The key risks associated with these three subsystems are shown in Figure 3.

AI in agriculture faces risks at the social subsystem level, including unemployment risks of AI replacing human labor, increased inequality due to socioeconomic barriers to AI adoption, and systemic exclusion caused by algorithmic bias (Table 2).

Overview of social risks of AI in agriculture.

Agriculture is a major field of employment in most countries (Raghuvanshi et al., 2022). With the widespread adoption of AI in agriculture, unmanned technologies are gradually replacing farmers, leading to significant social issues such as agricultural labor surplus and farmer unemployment (Leal Filho and Gbaguidi, 2024; Sandbrook, 2025). Currently, highly efficient unmanned technologies, including smart autonomous tractors, drones, and robots, are more and more important in agricultural production. These technologies are gradually replacing traditional agricultural labor, creating a threat to farmers’ livelihoods (Gonzalez-Rodriguez et al., 2024; Sparrow et al., 2021). Taking the field of plant protection and development as examples. An autonomous weeding system used in orchards can reduce reliance on manual labor (Han et al., 2025a), which may lead to a reduced demand for workers in orchards. Jobs such as plant protection technicians and plant quarantine inspectors are also facing the possibility of replacement by AI technologies like neural networks. More seriously, the unemployment issues caused by AI replacing human labor have already arisen in populist movements across Europe (Dorigo et al., 2025).

From the perspective of socio-economic stratification, AI in agriculture requires significant economic investment and professional knowledge related to AI. Not all agricultural producers can equally have access to and share in the benefits of AI development, which may increase social inequality (Bozeman et al., 2024; Lakhiar et al., 2025; Pimenow et al., 2025). At present, due to the relatively low digital literacy and income levels in some rural areas and among farming communities, especially smallholder farmers (Yuan and Sun, 2024), they face difficulties in accessing and applying modern AI technologies. This may lead to a widening wealth gap between large-scale and small-scale farmers (Khanna et al., 2024; Li et al., 2024; Maraveas et al., 2022). At the same time, this will further widen the gap in agricultural development between high-income and low-income countries (Kalimuthu et al., 2024), increasing global agricultural inequality.

In addition, farmers often lack sufficient understanding of AI clauses, placing them in a disadvantaged position of information asymmetry. This may lead to potential harm to their interests and widen the gap between different users (Fuentes-Penailillo et al., 2024; Ibrahim and Truby, 2023). AI in agriculture may also bring about issues such as widening wealth gaps and wealth monopolization between groups of different socioeconomic status, increasing socioeconomic inequality, and threatening social stability (Sandbrook, 2025; Sparrow et al., 2021). For example, farmers with better financial resources can fully deploy AI-powered plant protection equipment, thereby significantly improving pest and disease control, increasing irrigation efficiency, and reducing crop losses, whereas smallholders often cannot afford the high-cost AI equipment. This inequality in access to technology tends to widen the gap in crop yields between farmers of different scales.

It is worth noting that human society itself has certain biases, and these biases are easily copied into AI models. This leads to issues such as data bias and algorithmic bias, causing AI to potentially generate biased decisions that increase exclusion and undermine social fairness (Bozeman et al., 2024; Gardezi et al., 2024; Leal Filho and Gbaguidi, 2024; Li et al., 2024). For example, during the initial training phase, AI pest and disease detection models tend to rely primarily on data from mainstream crops, such as wheat and rice. As a result, data on diseases affecting non-mainstream crops like buckwheat and quinoa, which are grown by some farmers, may be overlooked. This means that farmers growing mainstream crops benefit more from AI pest detection models, while those growing non-mainstream crops may face systemic exclusion.

AI technology brings a range of positive values, such as precision, scientific, and high efficiency. Theoretically, these features are conducive to enhancing plant protection, optimizing agricultural decision-making, improving production output, and agricultural sustainability. However, AI is not omnipotent. Its application in real-world agricultural scenarios may face risks, such as uncertainty of technical devices, inaccurate decision-making, lack of decision traceability, and network security threats (Table 3).

Overview of technical risks of AI in agriculture.

AI in agriculture involves the integration of large AI models with a range of hardware devices, such as sensors, remote sensing, robotic arms, robots, and drones (Gardezi et al., 2024). Specifically, it relies heavily on hardware devices such as sensors and remote sensing to monitor and record various agricultural environmental data, including temperature, humidity, and soil fertility, then generates precise agricultural decisions through AI large models. However, devices such as sensors are exposed to agricultural environments with long-term humidity, intense sunlight, and rainfall. This causes significant wear and damage to electronic components and machinery. Research has summarized a series of data and case studies on device failures. For instance, sensors operating in muddy and humid farmland conditions may lead to inaccurate recordings of plant growth and crop health, while agricultural AI algorithms can produce errors during extreme weather fluctuations. Research shows that the error rate of GPS-based yield monitors is over 10% (Visser et al., 2021).

Consequently, AI technical devices in the agricultural environment inevitably face uncertainty risks, such as short circuits, hardware aging, internal malfunctions, and sudden failure, any of which can lead to the disruption of agricultural production (Galaz et al., 2021; Visser et al., 2021). For instance, sensors deployed to monitor plant growth and detect diseases may produce erroneous signals due to environmental interference such as waterlogging or insect attachment on the foliage, which can lead AI systems to misjudge plant health and subsequently generate inappropriate management recommendations. Currently, early warning alerts regarding plant pests and diseases generated by AI systems are primarily disseminated to farmers via SMS. However, signal instability and uncertainty in rural and agricultural environments often lead to message loss, preventing farmers from receiving early warning messages in a timely manner (A. Kumar and Patel, 2023).

Data training limitations bring risks of inaccurate decision-making. AI empowers agricultural production mainly by providing users with accurate agricultural decision-making. However, it is difficult for AI’s decision-making to maintain 100% accuracy, and its effectiveness is highly dependent on the completeness of the training data. The more comprehensive and extensive the training data is for the AI model, the more accurate the decisions it generates, and vice versa (Essenfelder et al., 2025; Mittal et al., 2024; Sparrow et al., 2021). Due to factors such as geographical complexity and changes in external agricultural environments, much agricultural data remains inaccessible. This frequently leads to challenges for AI models, including insufficient training data and poor data quality, reducing the accuracy of AI decision-making (Ali et al., 2024; Chen et al., 2021; Kpodo and Nejadhashemi, 2025).

When applying AI to agricultural production, the inaccuracy of decision-making always occurs. For example, the low resolution of Sentinel-2 satellite imagery makes it difficult to capture small-scale vineyard land data (Lacueva-Perez et al., 2025); training datasets for AI-based crop disease identification are often of poor quality and incompatible across platforms; AI systems struggle to learn from sufficient data, thereby compromising decision accuracy (Gupta and Tripathi, 2024); and an Indian machine learning model for rice production failed to incorporate the duration data of extreme weather events, leading to insufficient consideration during decision-making (Bowden et al., 2025). All of these factors bring risks of inaccurate decision-making. Additionally, AI models require continuous training on new data, so that models can optimize and upgrade their capabilities. Otherwise, they will generate decisions based on old data, failing to adapt to new planting scenarios and agricultural environments, leading to inaccurate decision-making (Kikon and Deka, 2022; Peters et al., 2020).

Heterogeneous agricultural contexts reduce the effectiveness of general-purpose AI models. Popular AI models on the market are often referred to as general-purpose AI, indicating that the same AI can frequently be applied across a wide range of different scenarios. However, specific AI models are often trained using agricultural data from particular regions, which may differ significantly from the real agricultural environments in other areas. It leads to general-purpose AI models failing to achieve 100% accuracy when applied elsewhere, often resulting in underfitting or overfitting issues that compromise the precision of agricultural decision-making (Anandhakrishnan and Jaisakthi, 2022; Gautron et al., 2022; Goyal et al., 2025).

It is essential that AI in agriculture gives careful consideration to regional differences, recognizing the limitations of AI models in diverse local agricultural contexts (Gautron et al., 2022; Hughes et al., 2022; Zewdu et al., 2025). For example, research has found that when neural network models with over 98% prediction accuracy are applied to new agricultural environments, they may generate erroneous decisions regarding planting, fertilization, irrigation, and other crop production practices (Kumar et al., 2022). General-purpose AI models trained in the agricultural context characterized by “few people and much land” will face limitations when applied to other regions characterized by “much people and few land,” as well as areas with different climate conditions (Quddus et al., 2025).

Many AI technologies, such as algorithms, machine learning, and deep learning, lack transparency during running, which is described as the black-box (Buyuktepe et al., 2025; Izquierdo-Bueno et al., 2024; Oikonomidis et al., 2023). AI users always find it difficult to understand and verify the validity of the outputs. More seriously, AI illusions and faked data have become increasingly significant issues, as some AI models potentially generate agricultural advice containing false information (Dorigo et al., 2025). Once AI models generate misleading decisions, users find it difficult to understand and identify potential errors. If there is a lack of professional review or the inability to identify issues based on one’s own expertise, it may lead users to make error decisions (Sandbrook, 2025; Stodle et al., 2025). In this context, it is easy to lead to technological runaway and risks of untraceable liability, thereby causing damage to user benefits. For instance, an AI-driven crop irrigation system suddenly delivered excessive water, causing root rot and reduced yields. However, due to the system’s lack of transparency, farmers were unable to determine whether the issue stemmed from sensor malfunction or algorithmic error, leaving them with no recourse for accountability.

AI in agriculture has reshaped farming models and planting methods, making agricultural producers highly reliant on technologies such as large-scale models, big data, and machine learning to conduct agricultural production. Operating in an open network environment, users may face risks of network attacks and privacy breaches (Bui et al., 2024; Kumar et al., 2021).

Network Attack Risks. The use of AI puts users in the digital world, where they manage agricultural production and plant management by operating various devices and monitoring data. This makes them easily face hacker attacks, highlighting the need to pay more attention to network security issues of AI in agriculture (Sparrow et al., 2021). Various agricultural technologies such as large-scale models, big data, agricultural robots, and autonomous tractors all require network support, making them likely to face threats like malware attacks and virus attacks during use (Hazrati et al., 2022; Khanna et al., 2024; Pastor et al., 2025). For example, a farm’s AI-powered plant management platform could be targeted by a cyberattack from a competitor, which could tamper with irrigation or fertilization instructions, leading to abnormal plant growth.

User Privacy Breach Risk. Data privacy and security are of great importance (Hazrati et al., 2022; Mmbando, 2025). However, highly data-dependent smart agriculture generates and records large-scale user data in real time, including user traces and sensitive crop data. Farmers lack strong awareness of network and data security. It can easily lead to data privacy breaches and information security risks (Balaska et al., 2023; Gao et al., 2024; Kumar et al., 2022).

Consequently, human activities have environmental impacts on the ecological system (Cote and Nightingale, 2012; Dorward, 2014; Yang and Sono, 2025). From the perspective of ecological subsystems (Table 4), the application of AI technologies may involve an over-pursuit of efficiency gains while ignoring environmental considerations, thereby causing unintended negative impacts on ecosystems (Arevalo-Royo et al., 2025; Galaz et al., 2021; Khanna et al., 2024; Yu et al., 2024).

Overview of ecological risks of AI in agriculture.

In theory, AI can help sustainable plant protection and biodiversity protection (Pandey and Pandey, 2023). As AI in agriculture is currently in its exploration and implementation phase, it may have some unintended negative impacts on biodiversity (Sajith et al., 2022; Sandbrook, 2025). For instance, in the same region, AI may provide similar farmers with homogeneous planting recommendations, leading to the problem of single crop cultivation and suppressing biodiversity (Leal Filho and Gbaguidi, 2024). AI-driven automated decision-making may fail to distinguish which species in a garden should be removed, potentially eliminating beneficial plants (Javed et al., 2025). Furthermore, AI technology can precisely destroy pests, but as mentioned above, AI technology faces risks of instability and inaccuracy. The complexity of farmland can limit the effectiveness of AI-based pest management (Javed et al., 2025). It may lead to excessive pest control, thereby damaging the biodiversity of agricultural fields and plant communities, which is contrary to the principles of sustainable plant protection and agricultural production.

Large-scale deployment and application of AI in agriculture may lead to large-scale, uncontrolled expansion of agriculture and livestock, leading to the shrinkage of plant communities such as forests and grasslands, and weakening the carbon sequestration capacity of plants, associated with a reduction in forest resources, increased carbon emissions, and impacts on the global climate. Research indicates that agricultural expansion driven by technological applications can increase deforestation and environmental decline, thereby bringing climate risks (Alam et al., 2023; Omotayo et al., 2025). For example, AI may tends to promote monoculture planting, leading to the clearing of vast swathes of forest, which causes native plant species to disappear and exacerbates global warming.

In addition, AI in agriculture will also increase energy consumption, leading to indirect impacts on the climate. AI technologies such as AI and robotics are known to have high energy demands (Licardo et al., 2024; Sandbrook, 2025). The excessive energy consumption associated with AI is accompanied by a rapid growth in electricity generation. High energy consumption, associated with rapid growth in electricity generation, may be accompanied by issues such as deforestation and increased carbon emissions that impact the climate.

Unlike traditional digital technologies, AI in agriculture primarily enhances efficiency by providing agricultural and planting decisions to farmers. However, as previously mentioned, the accuracy of AI decision-making can be influenced by various factors, and AI cannot guarantee absolute precision (Visser et al., 2021). This may lead AI models to generate wrong agricultural decisions, misleading farmers into adopting wrong agricultural practices, such as fertilization, irrigation, and pesticide application. The wrong decisions are not conducive to the ecological environment and plant protection, potentially causing pollution issues in the agricultural ecological system.

For instance, while AI could effectively reduce fertilizer use for agricultural producers, it sometimes generates incorrect and imprecise decisions for irrigation, fertilization, and other practices (Arevalo-Royo et al., 2025). In this context, users may adopt recommendations such as excessive fertilization and overuse of pesticides, leading to environmental pollution issues that directly harm plant health and reduce crop quality. Furthermore, in the application of AI in agriculture, various new materials, including nanomaterials, may be used (Ur Rahim et al., 2021). Although these new materials help improve efficiency, they may also cause some unintended soil pollution issues. Research indicates that nanoparticles and other “additives” used in modern, smart agriculture, may weaken soil microbial capabilities and reduce soil ecosystem functions (Xu et al., 2022), thereby threatening plant productivity and reducing agricultural output.