How AI Is Being Used to Predict Refugee Crises Before They Happen (2026)

The State of AI Powered Refugee Forecasting in 2026

The question of how AI is being used to predict refugee crises before they happen has moved from research papers into operational humanitarian systems. By 2026, UNHCR's Project Jetson, the Danish Refugee Council's Foresight model, the World Bank's High Frequency Phone Survey machine learning pipelines, and the JRC's Dynamic Data Hub all use statistical and machine learning techniques to anticipate displacement weeks to months in advance. The shift is significant. Five years ago, almost every humanitarian appeal was reactive. In 2026, a meaningful share of preparedness funding is being released against probabilistic forecasts rather than confirmed events.

The models do not predict the future in any deterministic sense. They produce probability distributions over plausible scenarios using inputs that include conflict event data from ACLED, food security indicators from IPC and FEWS NET, climate signals from NOAA and ECMWF, satellite imagery from Sentinel and commercial constellations, mobile phone metadata where ethically permissible, and social media signal analysis. The output is typically a range, for example a 60 to 80 percent likelihood that new cross border displacement from a given origin country will exceed a given threshold within the next 90 days. That is enough to pre position supplies and brief governments. It is not enough to declare a crisis.

What the Leading Systems Actually Do

UNHCR's Project Jetson, launched in Somalia and now expanded to multiple operations, uses gradient boosted decision trees and recurrent neural networks to forecast arrivals at Dollo Ado refugee camp in Ethiopia and similar sites. The Danish Refugee Council's Foresight platform produces three year displacement forecasts for more than 25 countries using ensemble models that combine conflict, governance, climate, and economic indicators. The IOM Displacement Tracking Matrix has integrated nowcasting models that flag anomalies in flow monitoring data within days rather than weeks. For a broader view of how these data systems compare, see Best Free Datasets for Tracking Global Displacement.

World Bank teams have used machine learning on satellite night light data and high frequency phone surveys to estimate population movements in areas where traditional registration has broken down. The results have been validated against UNHCR registration data in Sudan, Ukraine, and Venezuela, with mean absolute errors typically in the range of 10 to 25 percent at the subnational level. That accuracy is good enough to inform resource allocation but not precise enough to drive individual entitlements.

Where AI Forecasting Performs Well

The clearest wins are in protracted crises where the underlying drivers are well understood and the data signal is dense. Forecasting refugee arrivals at established camps in the Horn of Africa, anticipating seasonal pastoralist movements in the Sahel, and projecting cross border flows during predictable conflict escalation cycles are all tasks where current models outperform expert judgment alone. The models are particularly strong at integrating climate signals into otherwise conflict focused analyses, a domain where human analysts often lack the technical training to weight inputs correctly.

AI also performs well at anomaly detection. When call detail records, satellite imagery, and conflict event data all shift in the same direction within a short window, the models flag the anomaly faster than humans reviewing dashboards. That speed advantage matters in fast moving crises like the Sudan displacement crisis, where conditions can deteriorate from warning to mass movement in weeks.

Where AI Forecasting Struggles

The hardest case is the novel crisis. Models trained on historical data underperform when the underlying political or environmental dynamics shift in ways the training data did not capture. The October 2023 Israel Gaza escalation, the February 2022 full scale invasion of Ukraine, and the April 2023 outbreak of conflict in Sudan were all underweighted by leading forecasting systems in the weeks before they began. Human analysts caught some of the signals; the models did not.

The models also struggle with the political economy of displacement. They can forecast that conditions are deteriorating but cannot predict whether a government will close a border, whether a host country will permit registration, or whether donors will fund a response. Those decisions shape the observed displacement numbers as much as the underlying drivers do.

Ethical and Operational Constraints

Predicting displacement is not the same as helping displaced people. Several governance concerns shape what the 2026 generation of models can and cannot do. The use of mobile phone metadata is restricted in many jurisdictions and discouraged by data protection guidance from UNHCR and ICRC. Satellite imagery analysis of camps and informal settlements raises consent questions that the sector has not resolved. And forecasts that name specific origin and destination countries can be politically weaponized by host governments seeking to justify border closures.

The IASC has issued operational guidance requiring that AI forecasts be paired with human analyst review before being shared externally, that uncertainty be communicated explicitly, and that no individual level predictions be generated for protection sensitive populations. Compliance with that guidance varies across agencies.

The Trajectory

The direction of travel is clear. By the end of 2026, every major humanitarian agency will have at least one AI forecasting system in production, and the question will shift from whether to use AI to how to govern it. The most important development to watch is not model accuracy but integration. Forecasts that sit in a data team dashboard have no operational value. Forecasts that trigger pre positioned stock, anticipatory cash transfers, and diplomatic engagement are the ones that change outcomes for displaced people.

The honest read in mid 2026 is that AI has measurably improved early warning for established crisis types in well instrumented regions and has not yet meaningfully improved early warning for novel crises in data poor regions. Both findings will shape humanitarian budgets for the rest of the decade.

Sources and Further Reading

• UNHCR Project Jetson and Innovation Service: https://www.unhcr.org/innovation/
• Danish Refugee Council Foresight model: https://pro.drc.ngo/resources/news/foresight-displacement-forecasts/
• IOM Displacement Tracking Matrix (DTM): https://dtm.iom.int/
• ACLED conflict event data: https://acleddata.com/
• IPC food insecurity classification: https://www.ipcinfo.org/
• World Bank Data Blog on machine learning for displacement: https://blogs.worldbank.org/opendata
• IASC operational guidance on data responsibility: https://interagencystandingcommittee.org/

Figures and capabilities described above reflect publicly available documentation of these systems through mid 2026. Where ranges are given, they reflect validation results published by the agencies themselves rather than independent benchmarks.