人工智能与粮食安全还是数字不平等：量化农业生产力、社会包容、可持续生计与负责任土地管理之间的全球权衡

Abstract

Artificial intelligence (AI) is increasingly recognized as a powerful tool for addressing some of the most pressing challenges facing global agrifood systems, including food insecurity, climate change, resource scarcity, and the growing demand for sustainable agricultural production. However, an important question remains largely overlooked: do technological advances in AI translate into meaningful societal benefits? Drawing on a global evidence synthesis of 1,276 studies conducted across 112 countries, this review challenges the widespread assumption that technological success automatically leads to food-system transformation. We introduce the Artificial Intelligence Food Security Paradox Theory (AIFSP Theory), a novel conceptual framework suggesting that improvements in algorithmic performance, productivity, and operational efficiency often exceed corresponding gains in food security, livelihood resilience, social equity, and environmental sustainability. To examine this proposition, we combined systematic review methods with hierarchical Bayesian meta-analysis, network meta-analysis, structural equation modeling, explainable machine learning, and Monte Carlo simulations extending to 2050. We also developed the Artificial Intelligence–Food Security–Equity–Land Responsibility Index (AI-FSELR), a multidimensional framework designed to evaluate the societal value of agricultural AI beyond conventional technical performance metrics. The results reveal a consistent and previously underexplored pattern. While AI technologies significantly improved agricultural productivity (+24.8%), resource-use efficiency (+18.9%), and farm income (+13.8%), gains in food security (+8.6%) and social equity (+5.4%) were considerably smaller. Governance quality, digital accessibility, extension services, affordability, and farmer training emerged as stronger determinants of societal outcomes than algorithmic sophistication itself. Future scenarios indicate that inclusive and governance-driven AI pathways could increase societal benefits by more than 40% by 2050, whereas expanding digital inequalities may substantially limit these gains. These findings suggest that the future success of agricultural AI will depend not only on technological innovation but also on the institutional, social, and governance conditions that shape its adoption. The AIFSP Theory and AI-FSELR framework provide a new foundation for evaluating AI and guiding more equitable, resilient, and sustainable food-system transformations.

1 Introduction

The rapid emergence of artificial intelligence (AI) is profoundly transforming global agrifood systems and is increasingly recognized as one of the most influential technological developments of the twenty-first century. Recent advances in machine learning, deep learning, computer vision, remote sensing, robotics, digital twins, Internet of Things (IoT) technologies, edge computing, and intelligent decision-support systems are creating unprecedented opportunities to improve agricultural productivity, optimize resource use, enhance climate resilience, and support more adaptive and sustainable land-management strategies (Liakos et al., 2018; Kamilaris and Prenafeta-Boldú, 2018; Javaid et al., 2023; Paul, 2026). These innovations are enabling a transition from reactive and generalized management approaches toward data-driven, site-specific, and predictive agricultural systems capable of responding more effectively to increasingly complex environmental and socio-economic challenges.

The growing interest in AI-driven agriculture is occurring within a context of mounting global pressures on food systems. According to the United Nations, the global population is projected to reach approximately 9.7 billion people by 2050, significantly increasing the demand for food, water, energy, and natural resources (Gerland et al., 2022; Ritchie et al., 2022). Simultaneously, climate change, biodiversity loss, land degradation, water scarcity, geopolitical instability, and market volatility are intensifying pressures on agricultural production systems worldwide (IPCC, 2023; FAO et al., 2025). These interconnected challenges threaten not only agricultural productivity but also the resilience, sustainability, and equity of food systems, particularly in vulnerable regions already experiencing chronic food insecurity and environmental stress (FAO et al., 2025; Barrett et al., 2022).

Against this backdrop, artificial intelligence is increasingly presented as a critical component of future food security and sustainability strategies. AI-powered technologies are now deployed across virtually all segments of agrifood systems, including crop monitoring, disease detection, yield forecasting, precision irrigation, nutrient management, pest surveillance, livestock management, supply-chain optimization, market intelligence, and climate-risk assessment (Liakos et al., 2018; Zhou and Chen, 2023; Talaviya et al., 2020; Klerkx and Rose, 2020). Recent advances in computer vision illustrate the increasing technical sophistication of these applications. YOLO-based architectures have enabled integrated detection, segmentation, growth-stage classification, and real-time identification of horticultural crops, while specialized segmentation models have improved the detection of harvesting targets under challenging greenhouse conditions (Paul et al., 2024; Paul and Machavaram, 2024). Similar developments in cotton systems demonstrate the progression from fine-tuned YOLO-based phenological detection to graph-augmented object-detection frameworks for robust maturity assessment (Verma et al., 2024; Verma et al., 2026). The ability of AI to process large volumes of heterogeneous data and generate actionable insights in near real time has led many researchers and policymakers to view digital agriculture as a key enabler of sustainable intensification and climate-smart food production (Birner et al., 2021; Rose and Chilvers, 2018; Dong et al., 2021; Fan and Zhao, 2025; de Jong et al., 2021).

Reflecting this growing momentum, major international organizations—including the Food and Agriculture Organization (FAO), the World Bank, the Organization for Economic Co-operation and Development (OECD), CGIAR, UNESCO, and the United Nations—have increasingly emphasized the strategic role of artificial intelligence in accelerating progress toward multiple Sustainable Development Goals (SDGs), particularly Zero Hunger (SDG 2), Responsible Consumption and Production (SDG 12), Climate Action (SDG 13), and Life on Land (SDG 15) (FAO, 2022; FAO, 2025; Afrifa et al., 2026; OECD, 2019; Fuglie et al., 2020; Ben Abdallah et al., 2023; Harikrishnan et al., 2025; Camaréna, 2020; Marvin et al., 2022; Tagkopoulos et al., 2022; United Nations, 2024; Opola et al., 2025). These organizations argue that AI-enabled innovations can substantially improve decision-making, optimize resource use, reduce production risks, strengthen climate adaptation, support sustainable land management, and enhance the resilience of agrifood systems. Recent academic assessments further suggest that AI has the potential to facilitate broader food-system transformations through improved forecasting, precision management, environmental monitoring, and adaptive governance mechanisms (Harikrishnan et al., 2025; Kuhl, 2025; Agrawal et al., 2025; Iqbal et al., 2025; Wang et al., 2025; Kazanskiy et al., 2025). Consequently, artificial intelligence is increasingly viewed not merely as a technological innovation but as a strategic enabler of more sustainable, resilient, and inclusive agrifood futures.

Over the past decade, interest in agricultural AI has grown at an extraordinary pace. Bibliometric studies consistently show a sharp increase in publications addressing machine learning, precision agriculture, digital farming, and AI-enabled crop management since around 2015. This expansion reflects growing investments from governments, research organizations, agri-tech companies, and international development agencies. Today, AI applications span virtually every component of agricultural production and food systems, from crop disease diagnosis, yield prediction, irrigation scheduling, fertilizer optimization, and pest monitoring to livestock management, supply-chain optimization, market intelligence, and climate-risk assessment (Harikrishnan et al., 2025; Agrawal et al., 2025; El Jarroudi et al., 2024). Reviews by Liakos et al., Javaid et al., and Aijaz et al., and many others have documented substantial improvements in predictive performance, operational efficiency, and resource optimization across a wide range of agricultural contexts (Liakos et al., 2018; Javaid et al., 2023; Aijaz et al., 2025).

This technological expansion is also reshaping agricultural automation and Agriculture 4.0 architectures. Vision-guided robotic systems increasingly combine kinematic modeling with AI-based optimization to support autonomous harvesting operations (Paul and Machavaram, 2025), while decentralized blockchain architectures are being explored to strengthen traceability, data exchange, and coordination across agricultural value chains (Paul, 2026). In parallel, generative AI is emerging as a new component of Agriculture 4.0, with potential applications in knowledge support, decision assistance, and context-specific agricultural services, although important challenges related to integration, accessibility, reliability, and governance remain unresolved (Paul and Machavaram, 2025).

Despite these remarkable technological advances, a fundamental question remains surprisingly underexplored: to what extent do improvements in algorithmic performance translate into meaningful societal benefits? The vast majority of studies evaluating agricultural AI continue to focus primarily on technical indicators such as classification accuracy, precision, recall, F1-score, root mean square error, computational efficiency, or predictive performance (Liakos et al., 2018; Kamilaris and Prenafeta-Boldú, 2018; Javaid et al., 2023). While these metrics are indispensable for assessing the reliability and effectiveness of AI models, they provide only a partial and often incomplete representation of their real-world value.

Indeed, a disease-detection system capable of achieving near-perfect classification accuracy does not automatically improve food security, nor does an irrigation optimization platform that reduces water consumption necessarily enhance farmer livelihoods, strengthen social equity, or promote sustainable land stewardship. As emphasized by recent studies on digital agriculture and responsible innovation, technological efficiency alone cannot guarantee positive societal outcomes because the impacts of AI are mediated by a wide range of institutional, economic, social, and governance factors (Rose and Chilvers, 2018; Rose et al., 2021; Klerkx and Begemann, 2020; Rose, 2026; Ozor et al., 2024).

This challenge is becoming increasingly evident as AI technologies move from experimental settings into real-world agricultural systems. Emerging evidence suggests that the benefits generated by digital innovations are often unevenly distributed across regions, farming systems, and socio-economic groups. Large commercial farms generally possess greater access to digital infrastructure, technical expertise, financial resources, and data services than smallholder farmers, creating the risk that technological progress may reinforce rather than reduce existing inequalities (Birner et al., 2021; Fielke et al., 2020; Ingram and Maye, 2020; Stein, 2026; Abbasi et al., 2022; MacPherson et al., 2022; Klerkx et al., 2019; Mühl and de Oliveira, 2022; Lioutas and Charatsari, 2022; Lajoie-O'Malley et al., 2020; Sun et al., 2024; Abdulai et al., 2023; Eastwood et al., 2019). Similarly, improvements in productivity or resource-use efficiency do not necessarily translate into proportional gains in food security, resilience, or environmental sustainability, particularly when adoption barriers, governance limitations, or market constraints remain unresolved (Barrett et al., 2022).

A growing gap is emerging between technological success and societal