AI vs. UNHCR: Who Gets the Numbers Right on Global Displacement?

Framing the comparison honestly

The provocation in the headline is useful but misleading. AI does not compete with UNHCR for the global refugee count. UNHCR is the legal custodian of refugee data under the 1951 Convention; its numbers are the registration record. AI-derived estimates are something else: model outputs that estimate populations UNHCR cannot or has not yet counted. The right question is not who wins. It is where each method is the best available answer, and what the divergences between them tell us.

What UNHCR's numbers actually are

UNHCR publishes population statistics built from three layers: country-level registration data from proGres v4 and equivalent national systems, government-provided figures where UNHCR does not register directly, and estimates filed by partners for populations that fall outside formal systems. Each annual mid-year and end-year release is methodology-rich; the Refugee Data Finder exposes the underlying breakdowns by country of origin, country of asylum, and population type.

The strengths are well known. Registration data is individual-level, audited, and durable. The same record supports protection, assistance, and statistics. The weaknesses are also well known. People who do not register do not appear. Urban refugees who avoid the registration desk are systematically undercounted. Statelessness data is patchy because few states cooperate. And the global total lags real-world events by months because publication cadence is quarterly to annual.

What AI estimates actually are

The AI estimates that get attention in 2026 come from several sources. The World Bank has published machine-learning estimates of displaced populations using satellite night-light, mobile phone metadata, and high-frequency phone surveys. IOM DTM's nowcasting pipeline produces flow estimates from sensor and signal fusion. Academic teams have produced retrospective reconstructions of historical refugee flows using image classifiers on declassified imagery. And several commercial vendors sell satellite-derived camp population estimates.

The strengths are speed and reach. Models can produce a usable estimate within days of an event, for areas no registration system covers. The weaknesses are the absence of individual records (no model output can replace a registration), the dependence on training-data assumptions that may not hold in the area of interest, and the difficulty of expressing uncertainty in a way operational users will respect.

Where the methods diverge, and what that tells us

The interesting cases are the systematic divergences between AI-derived estimates and registration totals. Three patterns recur.

• AI tends to be higher than registration in early-crisis weeks. Satellite-detected tent clusters and SAR-detected population redistribution often outrun registration capacity by 20-60 percent in the first month. When registration catches up, the gap narrows. The lesson: the AI estimate was directionally right; the registration number lagged.
• AI tends to be lower than registration in protracted urban contexts. Models that count tents miss apartments. In Jordan, Türkiye, and Colombia, AI estimates of refugee presence consistently underread registered populations because urban refugees do not produce the spatial signatures the models are trained on.
• AI and registration both miss the same populations in closed environments. Where governments restrict humanitarian access, both methods degrade. Neither is a substitute for the other; both reflect the same blind spot.

The methodological literature treats these divergences as diagnostic, not adversarial. A 30 percent gap between a satellite estimate and a registration figure is information about which population is being measured and which is being missed.

What a good policy or newsroom workflow looks like

The newsrooms and policy units making the fewest unforced errors share a workflow.

• They report registration data as the legal-status anchor and AI estimates as the operational-presence range.
• They publish both numbers with the methodology distinction made explicit.
• They treat any reported gap as a methodology question to investigate, not a contradiction to adjudicate.
• They do not promote AI estimates over registration data in contexts where registration is the entitlement-defining record (protection cases, family reunification, durable-solutions tracking).
• They do not let registration cadence become an excuse for late reporting on events AI signals can pick up within days.

The honest takeaway

UNHCR's numbers and AI-derived estimates measure overlapping but distinct populations using methods with different strengths. The cases where AI clearly wins are early-crisis nowcasting and inaccessible-area estimation. The cases where UNHCR clearly wins are status determination, urban context, and any decision that touches an individual record. The cases where neither wins are closed information environments, and that gap is a policy problem, not a technology problem.

The takeaway most worth keeping: divergences between AI and registration are useful, not embarrassing. They are how the sector improves both.

Sources and further reading

• UNHCR Refugee Data Finder methodology notes: https://www.unhcr.org/refugee-statistics/methodology/
• World Bank Data Blog on satellite and phone-data displacement estimates: https://blogs.worldbank.org/opendata
• IOM Displacement Tracking Matrix methodology: https://dtm.iom.int/methodological-framework
• IDMC GRID methodology annex: https://www.internal-displacement.org/global-report/grid2024/
• Joint UNHCR-World Bank Joint Data Center: https://www.jointdatacenter.org/