How AI Is Being Used to Track Refugee Crises in Real Time — And Where It Falls Short

Real-time refugee tracking has finally arrived. It is uneven.

For most of the last two decades, the global picture of forced displacement updated at the speed of a UNHCR mid-year report. By 2026 that has changed. A combination of satellite revisit rates measured in hours, mobile-network signal aggregation, social-media geoparsing, and machine-learning models that ingest all three has produced something close to real-time refugee crisis tracking. Movements that once took weeks to confirm now show up in dashboards within days, sometimes within hours.

That capability is not evenly distributed. It works best where there is dense satellite coverage, cooperative telecom regulators, and a humanitarian operation already on the ground feeding ground-truth back into the models. It works worst where it is most needed: closed-border crises, urban displacement, and places where governments treat refugee counts as politically inconvenient.

The data sources doing the heavy lifting

Modern crisis tracking is multi-source by design. Five inputs dominate.

• Optical and SAR satellite imagery. Sentinel-2, Planet, Maxar and the Sentinel-1 SAR constellation give analysts sub-daily revisit over most of the world. ML models trained on labelled tent footprints, road compaction, and night-light anomalies flag new informal settlements and camp expansions without waiting for a human pass.
• Conflict event feeds. ACLED and the UCDP GED publish geocoded battles, explosions, and protest events on weekly or sub-weekly cadence. Models use spatial-temporal patterns to forecast the displacement that typically follows.
• Mobile network metadata. Where operators and regulators permit, aggregated and anonymised call-detail records or roaming signals provide a coarse map of population redistribution. This was used at scale during Ukraine 2022 and Sudan 2023.
• Border and registration systems. UNHCR proGres v4, IOM DTM flow monitoring, and national border databases remain the gold standard but lag by days to weeks.
• Social and news signals. Geoparsed Twitter/X, Telegram channels, and local news scraped at scale feed early-warning anomaly detectors.

The AI layer sits on top, fusing these into estimates of who moved where, when, and roughly how many. UNHCR Project Jetson, the Danish Refugee Council's Foresight platform, IOM's nowcasting pipeline, and the JRC's Dynamic Data Hub all use variants of this stack.

What real-time tracking actually delivers

Three operational wins are now visible across multiple crises.

1. Earlier supply pre-positioning. Models that flag arrivals 7-14 days in advance let agencies move blankets, water, and registration kits before a surge hits a border post. UNHCR has reported pre-positioning decisions in the Horn of Africa driven by Jetson forecasts since 2019; the practice is now standard. 2. Faster damage and camp-growth assessment. SAR-based change detection identifies destroyed neighbourhoods through cloud cover within 24-48 hours, shortening the window between event and assessed-needs estimate. 3. Tighter feedback loops between forecast and reality. Nowcasting models that compare predicted-to-observed arrivals catch model drift early, which matters when training data is months out of date.

For a reporter, the practical effect is that breaking-displacement stories can now be sourced against multiple independent signals on the same day, instead of waiting for an OCHA flash update.

Where it falls short

The honest list is longer than the wins list.

• Urban displacement is largely invisible. People moving into Khartoum suburbs, Aleppo apartment blocks, or São Paulo favelas do not generate the tent-cluster signature that ML models are trained to detect. Most urban IDPs are missed entirely.
• Closed information environments break the stack. When a government switches off mobile networks, restricts satellite tasking permits, or denies humanitarian access, three of the five data sources go dark at once. Models trained on dense data fail confidently in sparse data.
• Novel-crisis underprediction is the rule. Every major forecasting system underweighted October 2023 in Gaza, February 2022 in Ukraine, and April 2023 in Sudan in the immediate run-up. Models trained on past patterns do not see regime change.
• Double-counting and ghost cohorts. Fusing five data sources without rigorous deduplication produces inflated estimates. The 2024 round of inter-agency methodology reviews flagged this as the leading source of estimate disagreement.
• Protection risk is rising. The same systems that track displacement can be repurposed to track refugees. Vendor due diligence and access-control hygiene have not kept pace with capability.

What good practice looks like in 2026

The agencies doing this responsibly share a short checklist: every public estimate cites the primary registration source it triangulates against; uncertainty ranges are published, not hidden; protection-sensitive flows (origin and destination at named-locality level) are suppressed in public dashboards; and AI outputs are reviewed by a human analyst with operational knowledge of the corridor before they leave the building. The IASC operational guidance on data responsibility codifies most of this; compliance is uneven but improving.

For journalists, researchers, and policy teams, the takeaway is to treat real-time AI displacement tracking as a triangulation tool, not a source of truth. The most defensible reporting in 2026 cites at least two independent signals (for example a SAR-detected camp expansion plus an IOM DTM flow monitoring update) and an explicit uncertainty range. Anything tighter than that is overclaiming what the technology can do.

Sources and further reading

• UNHCR Innovation Service and Project Jetson: https://www.unhcr.org/innovation/
• IOM Displacement Tracking Matrix: https://dtm.iom.int/
• ACLED: https://acleddata.com/
• Copernicus Sentinel data hub: https://dataspace.copernicus.eu/
• IASC Operational Guidance on Data Responsibility in Humanitarian Action: https://interagencystandingcommittee.org/
• JRC Dynamic Data Hub: https://drmkc.jrc.ec.europa.eu/