Can Machine Learning See Maladaptation? A Systematic Evidence Map of AI in Global South Agricultural Adaptation to Climate Change

Abstract

Climate change threatens agriculture and rural livelihoods across the Global South, and machine learning (ML) is increasingly deployed to support climate adaptation. Nevertheless, the literature has not systematically mapped whether ML reaches beyond upstream hazard prediction into the downstream stages where adaptation outcomes and maladaptation become visible. This article develops a PRISMA-guided systematic evidence map of AI/ML in Global South agricultural adaptation. A Scopus search using a four-block Boolean string (ML/AI, climate hazard, adaptation/livelihood, and agrarian population) returned 310 records. Records were screened through three eligibility layers (topical relevance, climate-adaptation relevance, and Global South relevance) and were coded according to the deepest stage of the adaptation cycle reached. At the current evidence-mapping stage, 241 records were retained. The studies are dominated by publications from 2024-2026 and are concentrated in India, China, Pakistan, and Ethiopia. ML is heavily concentrated upstream: hazard prediction and monitoring (151 studies; 62.7%) and adaptation decision-making (82 studies; 34.0%). Downstream stages are almost empty: only 6 studies (2.5%) address vulnerability assessment, 2 studies (0.8%) evaluate adaptation outcomes, and none (0.0%) evaluate maladaptation. Within this mapped corpus, machine learning cannot yet see maladaptation: it predicts hazards but remains largely disconnected from the livelihood consequences of adaptation responses. Future research should shift ML from predictive accuracy toward causal, longitudinal, and justice-sensitive evaluation of adaptation outcomes.