IRC tests AI in crisis zones, finds one answer in four still fails

The IRC's argument lands in a sector under acute strain. The commentary notes there are more conflicts today than at any time since the Second World War, with almost 260 million people needing assistance and over 118 million forcibly displaced, while aid has contracted 25% since 2024 and clinics and schools are closing. Against that backdrop, the IRC frames AI not as a novelty but as a way to keep serving people when resources shrink — Signpost has reached 20 million people, and aprendIA is on track to reach nearly a million children in Nigeria by the end of 2026.

What distinguishes the IRC's position is the order of operations it describes. Before Signpost's AI assistant went near a client, it was run for six months across Greece, Italy, El Salvador and Kenya while moderators and protection officers scored answers on whether they were safe, client-centered and trauma-informed. The pass rate rose from 52% to 77%, but the IRC says it still failed roughly one answer in four — so the tool was not deployed on its own, every response went through a human, and the most sensitive questions never reached the model. The IRC says it publishes those failures because a sector that learns its lessons privately keeps repeating them.

The stakes the IRC describes hinge on whether that kind of discipline travels. The organization argues that because it works where the margin for error is smallest, applications like its own could help set guardrails for the whole industry, and that staying away would leave people in war, displacement and climate disaster as the last to benefit from tools everyone else enjoys. Whether other aid groups adopt the same publish-your-failures posture, or treat AI as something to avoid, is likely to shape how quickly these tools reach the people the IRC says need them most.