As famine data dries up, can AI step in?

The Trump administration’s steep cuts to foreign aid have hit the tools that humanitarian agencies and governments around the world rely on to forecast and stave off hunger crises before they spiral out of control.

The dismantling of the U.S. Agency for International Development has already shuttered the U.S.-funded Famine Early Warning Systems Network, or FEWS NET, the gold standard for forecasting food crises globally up to nine months in advance. Its collapse has also crippled the data-gathering capabilities of the Integrated Food Security Phase Classification, or IPC, a U.N.-backed tool used to measure the severity of food insecurity and declare famine.

These systems were the world’s best shot at sounding the alarm on famines — and they’re now barely holding on. With no obvious way to fill budget and resource gaps, the lost funding won’t come back anytime soon. That means artificial intelligence is increasingly in the spotlight.  

Researchers and humanitarian groups have been experimenting with machine learning and predictive analytics for years, aiming to make famine prediction faster, cheaper, and more scalable. The goal is to use technology to gather critical data — especially in places where conflict or cost makes on-the-ground assessments impossible.

Artificial intelligence, or AI, tools are already in use. Smartphone apps are helping small-scale farmers diagnose crop diseases, track weather, and get planting advice — even in places with limited internet or low literacy rates.

Now, that same energy is being channeled toward famine early warning systems, with a variety of AI-based hunger crisis prediction tools under development by the International Food Policy Research Institute, or IFPRI; the World Food Programme; the Intergovernmental Authority on Development, or IGAD; and various universities. 

IFPRI’s model could eventually assist humanitarians, policymakers, and development agencies once it is peer-reviewed and published, which could happen later this year.

“We are not trying to replace IPC or FEWS NET,” said Yanyan Liu, a senior researcher at IFPRI. “But we can say that our model, this method, is complementary,” said Liu, who earned a joint Ph.D. in Economics and Agricultural Economics from Michigan State University.

So how does it work? IFPRI’s model pulls in dozens of data points. These include food price data from the Food and Agriculture Organization and WFP, population statistics, soil pH levels, and satellite data on weather conditions and soil moisture levels. Critically, it also incorporates multiple indicators of conflict. The team uses data from the Armed Conflict Location & Event Data, or ACLED, project, including event counts for battles, explosions, and reported fatalities.

Conflict is one of the most important factors to consider when it comes to predicting hunger. The IFPRI team found that a 10% increase in conflict intensity corresponds with a 31% chance that people classified as “stressed” — or IPC Phase 2, according to IPC’s five-stage framework for assessing the severity of hunger crises — get pushed into Phase 3 or worse, which marks the threshold for humanitarian crisis. Stage 5 is famine.

Using all of these inputs, the model uses machine learning to forecast the extent and severity of acute food insecurity up to a year in advance. Validated against published IPC estimates, it accurately identifies 94.1% of cases classified as IPC Phase 3 or worse, Liu said. In addition, 77.5% of cases the model predicted to experience severe food insecurity materialized within three to 12 months.

The potential for these kinds of tools in conflict areas, where access to data is sparse at best, is especially promising.

“Our model can help fill in some gaps, some locations, in the conflict-affected setting, for example — where we cannot send people to go,” Liu said. But she’s quick to add: “We still need some field work in data collection. We need good data to keep improving our model.”

WFP and IGAD, with the support of Google, are working on a project designed to leverage the power of “machine learning to enhance early warning information systems.” Researchers at Georgia Tech and New York University are both working on something similar.

There is an “explosion” in the number of these tools, said Nicholas Haan, senior adviser for technology and food security at IPC whose team is tracking dozens of such efforts worldwide. That surge isn’t surprising, he added, given the rapid advances in AI and machine learning — and the growing need to do more with less as humanitarian crises intensify and funding tightens.

IPC isn’t just watching — it’s trying to harness the wave. The Trump administration’s dismantling of USAID has cut funding to many of the partner organizations that collect the data it relies on to issue its reports, forcing some to scale back or halt field surveys and weakening the evidence base for famine analysis.

So when it comes to AI, IPC is positioning itself as a global referee of sorts: validating and calibrating these new models so they align with the IPC’s common scale for measuring food insecurity. It’s a way to make sure that AI predictions are both useful and understandable.

“The risk is that models get things wrong,” Haan said. “The risk is that decision makers don’t understand the models. The risk is power goes to the people who own the models and run the models, not the people who are actually at the front line of decision making. But done right, each one of those risks can be dealt with.”

Another risk, Haan noted, is that the growing number of data sources and models could create a “cacophony” — too much information, and not enough clarity. That kind of noise, he said, ultimately overwhelms decision-makers.

But if IPC can serve as a platform to validate and calibrate these tools, it could make the data landscape more coherent — and more useful.

“We have the promise of making smarter decisions, which would save money, that cost less money to do because there’s less time for humans to be involved,” Haan said. “Like sending humans into the field to do field data collection is a very expensive endeavor. Not that that will go away. I still think it should happen, but it can happen in a much more streamlined way.”

Liu has been developing her famine prediction model for nearly a decade. But with traditional hunger data collection and analysis under financial strain, the stakes feel higher than ever.

“I feel the pressure,” she said. “I feel the motivation to do it.”