In a Rohingya camp in Bangladesh, Razia was wrongly struck from a food register. Her chatbot investigated and reinstated her within the day. Meanwhile, the elderly Azmat, unable to update his biometrics, was cut off without recourse. The technology fed neither refugee but routed Razia back to entitlement by escalating her problem to a human decision maker. Azmat could not access one in the automated welfare determination system.
This illustrates the promise and peril of saviourship by artificial intelligence amid the desperation spawned by the stark scale of global humanitarian retreat. OECD data reveals that official humanitarian aid fell 36 per cent to $15 billion last year, the lowest in a decade. Agencies have shed a third of their staff and closed hundreds of field offices. Meanwhile, the UN has given up universalism by hyper-prioritising 88 million out of the estimated 500 million people in dire need.
Humanitarian cuts could cost 500,000 lives annually with only half of them possibly salvaged through national mitigation efforts, once parallel reductions in development and health aid are also factored in.
Fast-growing AI capabilities have demonstrated efficiency yields of about 25 per cent in some industries, according to research done by the University of Pennsylvania and other such institutes. But similar hopes that AI could allow humanitarians to do more with less are misplaced – not because the technology is weak, but because it may be pointed towards the wrong problems.
To start, no algorithm manufactures a food ration or filters water. So, on the largest part of aid cuts – cash, food, medicine and other tangibles for human survival – AI offsets nothing.
However, AI can bring efficiency to the logistics and administration of aid delivery: assessing needs, organising transport, co-ordinating, complying, monitoring and reporting. This is perhaps 40 per cent of humanitarian effort – usually office-bound – while 60 per cent is irreducibly physical in camps, convoys, clinics and water points. The associated arithmetic of workflows points to the current generation of AI potentially substituting only a tenth of human effort.
That is hardly game changing because although AI can assist shrinking humanitarian capacities to function more efficiently, it cannot replace what is missing. Even such marginal utility is tempered by concerns. An overwhelming 90 per cent of resource-constrained field workers use whatever bots they can get, but less than a quarter of their employers have formalised AI policies and safeguards. At a time when AI capabilities are doubling every seven months, according to Metr, an independent research organisation that evaluates frontier AI models, the risks of harm-causing errors are significant.
This is not an argument for throwing the rapidly growing AI baby out of the humanitarian bath water but for redoubled care when walking with it. The benefits are strongest in anticipatory action.
During the Los Angeles fires of January 2025, predictive AI modelling assessed fire trajectories, optimised evacuation routes and prioritised firefighter deployments. In Bangladesh, flood predictions modelled potential livelihood effects to trigger household cash transfers before the waters arrived. The emergency response mechanisms of the UN, World Bank and the International Red Cross and Red Crescent Movement are activated more quickly nowadays when certain thresholds are crossed.
In Africa, where previous delayed Ebola responses caused serious human and economic consequences, AI-powered disease surveillance platforms predict hotspots and reduce response times. AI logistics engines in DR Congo anticipate population movements and pre-position aid.
However, despite the considerable life-saving potential of semi-automated anticipatory action, less than 0.5 per cent of funding goes there, according to Devinit, an independent developing and consulting organisation. More cruelly, the first casualties of aid cuts have been long-standing food security and health-monitoring systems that produce the large-scale data on which AI trains.
We need humanitarian-minded tech bros to invest the billions needed to underpin greater machine learning for more reliable humanitarian AI. Imagine the consequences for a Darfuri family sent by a hallucinating bot along a road with waiting armed predators, only to reach an empty distribution centre.
This illustrates how life-saving AI in benign disaster contexts becomes a dual-use weapon in wrong hands. “Do no harm” is a sacred humanitarian mantra but keeping secure the data of vulnerable people in chaotic emergencies is challenging. Data leaks show that the substrate that feeds AI to enable smart aid can expose hapless recipients to terrible abuses.
Thus, the AI directing refugees towards safety can allow belligerents to pre-block evacuation routes and humanitarian convoys or warn a neighbouring country to close borders early. At a time when international humanitarian law is observed more through violations, AI-driven protection could, perversely, enhance risk.
But could AI be a fairer aid dispenser in a system discredited by diminishing neutrality and impartiality because polarised geopolitics are compromising humanitarian choices? Could an un-sentimental, un-feeling AI be more consistent and trustworthy than fallible and emotional humans?
No, because human prejudices already baked into AI models cannot adjudicate between contested values. Thus, “life-saving” – the bedrock for rationing scarce aid – splinters into incompatible meanings around who to serve and how. Choosing is a moral act with all its risks of bias. This is impossible to code into an algorithm.
A deeper danger lurks with AI trained on historical data. It faithfully learns that forgotten crises are low priority because they were forgotten, thereby entrenching yesterday’s neglect into tomorrow’s logic. Besides, an AI appearing to make triage fairer chooses the contested factors that are quantifiable and not necessarily humanly meaningful. That is not fairness, but values-imposition disguised as technocratic objectivity.
Traditional humanitarians proudly walk the last mile to ensure that that the most vulnerable are not overlooked. AI’s probability-based protocols cannot do that, and so gaps widen when the digitally disconnected, remote or incapable get cleansed out of databases.
In fairness, flesh-and-blood aid workers already have a kind of AI run by hand, with severity scores and spreadsheets converting anguish into services that grow sterile with metrics and value-for-money considerations. Thus, AI did not begin any dehumanisation but threatens to complete it by making computations invisible. At least a human running the cruel arithmetic feels the burden of its weight. An algorithm triaging millions does so without itself hurting. So, an exhausted, impoverished sector may be sorely tempted to let the machine carry a burden that should remain humanly unbearable.
When the capacity to be troubled by a choice that affects vulnerable people becomes, instead, an inefficiency to be optimised, the conscience of the humanitarian enterprise is lost. Hence, the direction in which we point AI at is an immensely consequential choice.
Turned towards humanitarian bureaucracies to make rationing frictionless and responsibility diffuse or unaccountable could allow AI to finish hollowing out the very humanity the word enshrines.
But if AI clears the desk of distractions to make space for the humanitarian to sit in the same room as the beneficiary and help them to challenge and navigate life’s adversities, then AI can be a powerful agent to restore something that our humanitarian system mislaid long ago. That is the dignity which is, after all, largely a matter of who gets to decide on what, regardless of how dire the attending circumstances are.
That is the true AI for humanitarian good, a project still in its very early stages.
