This year has seen the marking of a grim milestone, with two famines recorded simultaneously in Gaza and Sudan, collectively impacting millions. It’s the first time that this has happened since the universal food insecurity measurement system, known as the Integrated Food Security Phase Classification (IPC) was established in 2004.
While ceasefire in Gaza has led to hope that food security in the territory will now improve, humanitarian situations elsewhere in the world are escalating due challenges including humanitarian aid cuts. A report published by the World Food Programme (WFP) this week highlighted how key challenges include record levels of food insecurity in the Democratic Republic of the Congo; soaring malnutrition rates in Afghanistan, where food assistance is now only reaching 10 per cent of the population; and a spiraling crisis in Haiti, where food aid recipients are now receiving receiving food worth half of WFP’s standard monthly rations as a result of budget constraints.
“The world is facing a rising tide of acute hunger that threatens millions of the most vulnerable – and the funds needed to help us respond are drying up,” said WFP executive director Cindy McCain.
While “IPC 5” is seen as constituting a famine, even in in IPC 3 or 4 – which last year impacted 300 million people around the world – there is a risk of children dying from starvation, according to Save the Children. As Simon Levine, famine expert at think tank ODI Global, explains: “IPC 2 is really not very good at al, three is when you start to get worried, four is when you’re very worried, and five should never ever happen.”
But the ability for humanitarian organisations to detect and monitor food insecurity on the IPC scale was thrown into serious question this year when US-government funded the Famine Early Warning System Network, or FEWS NET, widely seen as the gold standard tool in famine detection, went dark, after Trump’s sudden closure of the US Agency for International Development (USAID) left it unfunded in January.
First established in 1985 after famine in Ethiopia gripped the world’s attention, FEWS NET has been producing monthly reports on a large number of countries through its tried and tested analytical framework. It has been honed through forty years of experience working with difficult data to reach food insecurity assessments that are compatible with the IPC scale – encompassing everything from harvest times to hazard risks to prices available at local markets.
“We build scenarios that are rooted in how people live,” says Tim Hoffine, deputy chief of party at FEWS NET. “As analysts, we integrate data and evidence, but it’s not based on an algorithm or a single survey.”
Much to the relief of NGOs and governments, FEWS NET relaunched in June. But there has been a clamour of voices asking, in a world with less overseas aid available, whether algorithmic or AI-based models can provide cheaper, smarter alternatives than a 40-year-old system.
“We are all feeling this pressure at the moment from governments and agencies saying: ‘Why don’t you just use AI for everything?’,” Dave English, principal engineer at FEWS NET, tells The Independent. “We have also been approached by people that have built these huge composite indicators that are trying to predict the whole thing in one go.”
Organisations like the International Food Policy Research Institute are developing AI tools to predict famine, while government departments focused on aid are pouring money into machine learning processes: The UK Foreign Office, for example, is set to launch two major AI for development initiatives at the upcoming G20 meeting in South Africa.
But despite engaging with such people “as much as we can”, FEWS NET is yet to see anything deemed “particularly useful” in the world of AI-driven famine detection, says English. There are a number of reasons why this is the case, according to experts.
Clear benchmarks but difficult data
“No famine”, wrote Amartya Sen, the world’s leading expert in famine theory in 1999, “has ever taken place in the history of the world in a functioning democracy”. These events are not about there being a lack of food, Sen argues, but about political dynamics preventing food reaching people.
Such political dynamics also filter down to the quality of available data, which usually requires experience and consideration to work around. “A lot of data is out-of-date, inadequate, or uncertain, which means that rating food insecurity becomes a consensus-based process between multiple experts,” says ODI Global’s Levine.
“Governments can manipulate data: crop yields can be inflated in national statistics, or things like the reported inflation rate can be wildly different to what people are experiencing on the ground,” adds FEWS NET’s Dave English.
While on paper, there are clear benchmarks of what constitutes IPC 5 – including at least 20 per cent of the population need to be facing extreme food shortages; 30 per cent of children are acutely malnourished; and two out of 10,000 people are dying because of hunger – in practice, working out whether a situation meets those thresholds involves experts weighing up problematic data and political trade-offs, to be able to complete their overall assessment.
AI struggles to replicate the role of the human brain into the careful process of collecting and assessing such data: “Just taking data that is there and throwing it into a machine learning model will provide wildly inaccurate results that don’t really show you what’s what’s going on in a region,” says English.
Even if you do build a workable machine learning model around famine, there’s always a chance that a shock that has not been seen before will come along, which it cannot recognise, English adds. When Russia invaded Ukraine in 2022, for example, there was suddenly a need to account for the risk that one of the world’s main exporters of grain might no longer be connected to the global market, which posed a major, but atypical, threat to food security in Sub-Saharan Africa, which an AI model would likely not pick up.
Specific situations can throw up all kinds of scenarios that AI would struggle with. English cites another example where it appeared from satellite monitoring that herds of animals in a certain region had all died – but in actual fact, the satellite monitoring was only taking place at midday, which is when the herds shelter under trees for shade and were unable to be seen, but very much still there.
Inbal Becker-Reshef, managing director at Microsoft AI for Good Lab, points out that machine learning tends to lend itself to large amounts of data being available – but determining food insecurity and agriculture production depends on “a very local context”.
Often FEWS NET analysts carry out carefully considered, assumption-based modelling – for example, comparing one area of a conflict that has better data with an area that is data poor – which is something that AI would struggle with, says Tim Hoffine.
Experts FEWS NET spoken to by The Independent stress that they are not ideologically opposed to using AI – and indeed, the organsiation is increasingly using machine learning to boost some of its functions, including in AI-based weather forecasting models, as well as in agricultural and population modelling.
“It’s not about replacing FEWS NET with AI, but about how we can use it in a responsible and effective way,” says Weston Anderson, an agroclimatologist working at FEWS NET. “We need to maintain scientific expertise and on the ground expertise – but great if we can use AI to free up more time for analysts to do their job effectively.”
As AI models improve there may well be an ever greater role to play in functions related to famine monitoring. But were that to happen, there must be absolute certainty in its capabilities, the stakes for things going wrong are simply too high.
“The real marker of our success is that we never wake up today and are surprised to see a famine,” says Tim Hoffine. “The risk if we do start to rely on a new system and it goes wrong is that lives will be lost, which could have been avoided.”
This article was produced as part of The Independent’s Rethinking Global Aid project
