Can AI End Hunger? What Puts Food on the Table, and What Only Makes Headlines

THE NUMBER THAT MATTERS. The latest SOFI report from the FAO and four other UN agencies puts chronic hunger at 645 million people in 2025. That is 7.8% of the world's population and the third straight year of decline. The progress is uneven. For the first time, Africa has more hungry people than Asia: about 309 million versus 292 million. In Africa, one person in five goes hungry. The 2026 Global Report on Food Crises shows the harshest side of the problem. In 2025, 266 million people in 47 countries faced acute food insecurity, meaning a lack of food that threatens lives or livelihoods. That year, famine was confirmed in parts of Gaza and Sudan. It is the first time in the report's ten-year history that two famines were confirmed in the same year. The report names armed conflict as the main cause, followed by drought and aid cuts. Our thesis: AI already helps farmers grow more and lose fewer harvests, but it does little against the worst hunger, which comes from war, poverty and political will. Confusing the two is the mistake behind most of the headlines.

WHAT AMARTYA SEN ALREADY EXPLAINED. In 1981, the economist Amartya Sen, who later won the Nobel Prize, published ‘Poverty and Famines’. His main case was the 1943 Bengal famine, which killed between 1.5 and 3 million people, depending on the source. Sen found that food output that year had fallen from the year before but was higher than in other years without famine. People did not die because rice ran out. They died because the war drove prices up and rural wages did not keep pace. Sen called this the ‘entitlement’ approach: you go hungry when you lack the means to get food, even if the food exists. His thesis has critics, but its core idea has held up for four decades. Our reading: no AI model opens a humanitarian corridor or raises a day labourer's wage. When a tech executive promises that AI will end hunger, it is worth asking which hunger they mean. If they mean the small farmer who loses a harvest, AI can help a great deal. If they mean Sudan, the problem lies elsewhere.

WHAT ALREADY WORKS. The strongest progress is in weather forecasting. In July 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) put AIFS-ENS into service. It is the centre's AI ensemble system: 51 separate forecasts that, together, estimate how likely each scenario is. According to the ECMWF, it runs more than 10 times faster than the centre's physics-based model and uses about 1,000 times less energy. It is not better at everything. A 2026 preprint finds that the classic physics-based model still forecasts wind more accurately. That is why the ECMWF runs both systems side by side. The most interesting real-world use came in India. In summer 2025, the Ministry of Agriculture sent text messages forecasting the monsoon's arrival to about 38.8 million farmers in 13 states. The forecasts blended NeuralGCM, an open model from Google; Europe's AIFS; and 125 years of rainfall data. According to Google and the University of Chicago, they predicted the monsoon's arrival up to a month ahead. For someone deciding when to sow, that month can be worth a harvest. The second front is diagnosing crop diseases by phone. The PlantVillage Nuru app identifies cassava diseases from photos of the leaves, and it works without a network signal. In a peer-reviewed study in Tanzania, it was right 65% of the time. Farm advisers were right 40% to 58% of the time, and farmers 18% to 31%. When the app checks six leaves per plant, its accuracy rises to 74–88%, roughly the level of an expert. The third front is early warning, such as FEWS NET, the US famine warning system. Here the lesson is not technical. According to Devex and NPR, the system was offline for about six months in 2025 because of the US freeze on foreign aid. In 2026, it went weeks without updating countries such as Somalia, Afghanistan and Yemen. A warning system only works if someone pays for it, and then someone acts when the alarm goes off.

WHAT IS PROMISING AND WHAT IS HYPE. Farm advice from generative AI is the most appealing promise. Farmer.Chat, from the nonprofit Digital Green, answers questions over WhatsApp or by voice in local languages. According to the organisation's own estimate, cited by OpenAI, it could cut the cost of farm advice from $35 per farmer to $0.35. That is a projection, not a measured result. One warning sign is worth remembering. In a randomised trial in Kenya, researchers from Berkeley, Harvard and other universities gave hundreds of small business owners a GPT-4 mentor on WhatsApp. On average, it made no difference. Those already doing well improved by 15% to 20%. Those doing poorly got 8% to 10% worse. According to the authors, the weaker group asked about hard problems, such as drought or lack of capital, and followed generic advice, such as cutting prices. It is the same pattern we see in healthcare: AI works best with human oversight and training alongside it. Weeding robots are more mature, but their economics are not yet proven. Carbon Robotics' LaserWeeder uses computer vision to find weeds and burns them with lasers. An independent 2024 study confirmed that laser weeding removes at least 97% of weed biomass, about as well as herbicides. The company says it cuts weed control costs by up to 80%. We found no independent audit that confirms this once the machine's price is counted. And today it pays off on high-value vegetables, not on the maize fields that feed most people. AI-driven gene editing is the longest-term bet. The company Inari combines multiplex editing, which changes several genes at once, with models that predict which edits will work. It has raised about $720 million. Its goal is to raise soybean, maize and wheat yields by 10% to 20% without using more inputs. For now, that is a goal, not a verified field result. The hype is in vertical farms: indoor crops grown on stacked shelves under artificial light. Plenty filed for bankruptcy in March 2025 after raising nearly $1 billion from investors including SoftBank, Walmart and Jeff Bezos. Bowery shut down in November 2024 after raising hundreds of millions. AeroFarms and AppHarvest filed for bankruptcy protection in 2023. The reason is simple: growing with electric light costs more than growing with sunlight. They also grew mainly lettuce and strawberries, which provide few calories. They were never a solution to hunger.

OUR READING AND THE IMPLICATIONS. In the short term, we see three real risks. The first is the gap. If AI advice mainly helps farmers who already have land, credit and training, it may widen inequality in farming instead of narrowing it. The Kenya trial points that way. The second is concentration. If edited seeds are patented and field data stays on the platforms of a few machinery makers, the value will flow to them and not to small farmers. The third is political fragility. The best warning tools depend on budgets that can be cut by decree, as the FEWS NET shutdown showed. In the long term, we are optimistic, with good reason. Climate change will make rainfall more erratic, and AI already offers cheaper, more accurate forecasts exactly where they are needed most. Seeds designed to resist drought and pests could protect harvests in the coming decades. And the cost of expert advice is falling towards almost zero. That is the abundance we argue for: fewer hours of exhausting uncertainty on the farm and more room for each person to choose their work. But there is a condition. What already works rests on open public goods: Europe's AIFS model, the open-source NeuralGCM and India's historical rainfall data. It is the same case we make for open source: it democratises. Our proposal for measuring progress is simple. Forget how many farmers download an app. Measure how many harvests are saved, how much the poorest farmers' incomes rise, and whether warnings arrive in time and get acted on. AI can help the world grow more than enough food. Getting that food to the people who need it still depends on peace and politics.