I’m a humanitarian leader who’s all in on AI. Here’s why

AI can reinforce bias. It can hallucinate. It can expose sensitive information. It is used geopolitically in dangerous ways, with applications that make war easier and democracy harder. We are even now, more than ever, grappling with a conversation about AI’s existential risks. 

So people are surprised when I tell them the International Rescue Committee, where I lead research and innovation, is building AI into how we serve people living amidst conflict and crisis. 

And after years of testing AI in refugee camps, conflict zones and disaster responses, I’ve reached a conclusion that surprises them even more: Humanitarian organizations shouldn’t only be among AI’s biggest skeptics. We should also be among its strongest advocates; its frontier users; and its most active learners. We need to be all in on AI. And we need to lead, not follow, on mitigating AI’s risks. 

Here’s why, and how. 

This moment demands invention… 

This is a terrible moment for humanitarian aid. There are more conflicts today than at any time since the Second World War, and climate disasters are accelerating. Almost 260 million people need assistance, including over 118 million forcibly displaced — a grim world record — while governments retreat from funding the response. Aid has contracted 25% since 2024. Clinics are closing, schools are shutting their doors, vaccines are not being delivered. Hundreds of thousands of entirely preventable deaths, if not millions, are expected between now and 2030.

That is the moral case for AI in humanitarian settings. It is also why we cannot afford to get it wrong. People in crisis have the least power to refuse a technology, contest its errors, or walk away from it. They should never be anyone’s test case. So the question was never whether AI could help, but what it takes to find out safely.

In our sector, the benchmark that matters is impact. Signpost, our digital information service, has reached 20 million people navigating the hardest moments of their lives, nearly 591,000 of them through one-to-one guidance from moderators, many displaced themselves. An AI mentor called aprendIA reaches teachers through WhatsApp, because that’s the technology they already have. In Nigeria it is on track to reach nearly a million children by the end of 2026 and 6.9 million by 2028, at an estimated $0.20 per child.

We’re piloting the use of AI and satellite imagery to find children who never appeared in  census records as conflict moves entire communities. If you don’t know where children are, you can’t vaccinate them. By adapting delivery routes as climate and conflict conditions change, the tool could cut cost per child by 25% while surfacing 10 to 20% more children. Pilots launch in Somalia next quarter.

Globally, 43 million children under five suffer acute malnutrition. Our simplified treatment programs work, and are 20% cheaper than the standard treatment mechanisms; but manual, error-prone data collection holds them back. So we’re testing photo-to-digital capture to replace hand transcription, flagging children who drop out of care, and reading rainfall and displacement signals to act before caseloads surge.

Again and again, AI, even with its flaws, has the potential to deliver, enabling us to reach more people in need, faster, for less. But what about its potential for harm? 

… it also demands proof

Humanitarians, like doctors, are guided by the principle “do no harm”. When mistakes can cost lives, “move fast and break things” isn’t a strategy. So we built a prototyping and evaluation capability whose job is not to launch AI but to break it first, creating a resource the whole organization draws on.

Here is what that looks like. 

Before Signpost’s AI assistant went near a client, we ran it for six months across Greece, Italy, El Salvador and Kenya while moderators and protection officers scored its answers against three standards: is it safe, is it client-centered, is it trauma-informed. The system improved substantially, as its pass rate climbed from 52% to 77%, and staff reported working about 70% faster. 

But it still failed roughly one answer in four.

So we did not deploy it on its own. Every response went through a human, and the most sensitive questions never reached the model at all. That is the part of responsible AI that doesn’t make headlines: sometimes the finding is that a tool isn’t good enough yet, and the discipline is to say so and stop. We publish those findings, failures included, because a sector that learns its lessons privately keeps repeating them.

The same discipline governs the rest of the work. We track guidance from data protection authorities and cybersecurity agencies, sit in peer networks where security professionals share emerging threats in real time, review use cases carefully, and revisit our own policies continuously, because what was sufficient a year ago may not be today. We design for the constraints we actually face — low bandwidth, old phones, dozens of languages — and build nothing a human can’t audit, override or switch off. AI cannot replace a caseworker with lived experience, and we aren’t trying to. The goal was never to replace people, but to let teams reach far more of them.

Responsible AI isn’t a slogan. It’s operational discipline. And because we work where the stakes are highest and the margin for error smallest, applications like ours could help set guardrails for the whole industry. We have stronger incentives than almost anyone to use AI carefully,  and to show our work.

Staying away would mean accepting that people living through war, displacement and climate disaster become the last to benefit from technologies everyone else enjoys. It would mean telling families whose aid has just been cut that we’re declining to explore tools that could change their lives — leaving “AI for good” on the table.

I don’t think that’s an ethical position.

The humanitarian sector has spent decades solving some of the hardest operational problems on Earth. If AI can help us solve them better — carefully with the right guardrails, transparently, and with evidence to prove it — we shouldn’t apologize for using it. We should insist on it.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com