AI and Food Security: Can It See Hunger Before It Strikes?

In 2022 the world wasted 1.05 billion tonnes of food (including parts that are inedible), 60% of it in households, 28% in food services and 12% in retail, according to the UN Environment Programme. In the same year about 8.7% of the world's population was going hungry, according to the 2025 UN report. The two figures belong to roughly one year, so neither is explained by an absolute shortage of food; the cause is the road food travels from field to table, along which tonnes and lives are lost.

Hunger is a question of access

The UN report The State of Food Security and Nutrition in the World 2025 estimated that about 673 million people (8.2%) experienced hunger in 2024, within an estimated range of 638 to 720 million. That is down from 8.5% in 2023, but hunger rose in most subregions of Africa and western Asia, and in Africa it exceeds 20% of the population (about 307 million people). The agencies behind the report project that about 512 million people will still be chronically undernourished in 2030; that is a projection, not a settled verdict.

The economist Amartya Sen argued in Poverty and Famines (1981) that famines arise mostly when groups of people lose their ability to obtain food, through work, purchase or exchange, even when food is available in aggregate. His example was the Bengal famine of 1943, when supply was not much lower than in years without famine while prices rose faster than wages. Later researchers have disputed his reading of the supply figures, so the matter is debated, though the influence of his idea remains great.

What can the algorithm do?

Machine learning's strength here is early warning. In a World Bank working paper (September 2020), a Random Forest model forecast food crises monthly at district level, drawing on rainfall, vegetation indices, conflict and food price inflation, and was compared with the standard phased food-security classifications (phase 3 or worse) recorded for 1,162 districts in 21 developing countries between 2009 and 2020. The official near-term outlooks had missed about 74% of actual crises, and the model's weighted error fell from about 26% to about 11% at horizons of four and eight months.

The figure must be read within the limits its authors state: a working paper its authors describe as work in progress, gaps in the data, validation based on a random split of data that are linked in time and space, a short time-based test that fell in a period of severe crises, and large errors in some countries such as Somalia and the Democratic Republic of Congo. Even the error weights are a policy decision, because the question is which is worse: a false alarm or a crisis missed.

Where food is lost

The Food and Agriculture Organization estimates that about 13.3% of world production is lost between harvest and the retail stage (about 1.31 billion tonnes), and after that the UN Environment Programme puts waste at nearly a fifth of the food available to consumers. Better forecasting of demand, transport and storage might be expected to reduce such losses, but we found no documented measurement on which to base a figure for what algorithms achieve here, so we state none. What is clearer is that 60% of waste occurs in homes, a matter for good management, thrift and protecting God's bounty, which no algorithm can take over.

The model's limits

A model trained on data from other regions may misread our soil, our rainfall and our habits; this is a general principle of machine learning, not a result measured here. Groups whose data are not collected do not appear in the model, so their representation in the data becomes a matter of justice. A model is also likely to weaken when the climate departs from what it learnt from the past, and the farmer's intuition and the experience of field workers remain partners that cannot be dispensed with.

What to do

1. Pair early warning with early response; a forecast with no action feeds no hungry person.
2. Invest in reliable local data before buying algorithms.
3. Make cutting waste in homes, restaurants and markets a priority alongside raising output.
4. Keep the final decision with a person, and ask every model to state the limits of its accuracy.

Provision is in God's hand, and ours is to strive, take the means and be fair in sharing. If hunger becomes predictable, there is no excuse left for failing to prevent it.