Beyond Prediction: Data, Baselines, Explanation, and Causation in Machine Learning for Food Insecurity

ABSTRACT More than 280 million people across over 50 countries face acute food insecurity, yet the evidence systems that classify severity and trigger assistance remain slow, costly, and geographically uneven. In the past 10 years, the use of artificial intelligence (AI) and machine learning (ML) has been suggested as a cure‐all and has already been applied to the prediction of household welfare, to nowcast food consumption indicators at the subnational level, to forecast Integrated Food Security Phase Classification (IPC) transitions, and to the use of satellite and telecom proxies. This review examines that literature and is undertaken with three aims. First, it records publicly available global and regional datasets, their coverage, latency, access routes, and their caveats in the form of modeling. Second, it resets the benchmarking standard, correcting the common misconception that tree ensembles are more data‐hungry than deep learning, and requires comparison to persistence and climatology and conventional statistical baselines under spatially and temporally blocked validation. Third, it does not just stop at prediction but looks at post hoc explanation, causal machine learning, and alignment with each IPC dimension before going on to consider label scarcity, algorithmic bias, and the digital divide.