AI Reconstructs 34 Years of Global Migration in Unprecedented Detail

Understanding how and why populations move across the globe is crucial for shaping social policy, managing economic shifts, and directing humanitarian aid, yet historical records remain surprisingly fragmented. To fill these massive blind spots and correct a longstanding bias toward Western data, a new AI-driven approach has reconstructed annual migration flows across the world, scientists report today in Nature. The work, led by senior author Guy J. Abel at the University of Hong Kong, with first author Thomas Gaskin, utilized an ensemble of deep recurrent neural networks to generate a globally consistent, high-resolution map of human mobility.

The research team integrated a vast array of fragmented resources, merging traditional demographic stocks, official border statistics, and recent digital traces from platforms like Facebook. To capture the complex drivers and "memory" of human movement, the neural networks were informed by geographic, economic, and political variables, such as gross domestic product, life expectancy, and conflict data. This unified modeling framework successfully mapped annual origin-to-destination migration for 230 countries and regions spanning from 1990 to 2023.

The resulting dataset reveals that global migration has surged significantly this century, climbing from approximately 13 million annual movements in 2000 to around 35 million in 2023. This upward trend outpaces global population growth, with per-capita migration more than doubling over the same timeframe. The deep learning model successfully captured sustained regional shifts—such as the estimated 19 million movements from South Asia to Gulf states since 2010—while also pinpointing acute historical shocks, like the 1994 exodus of nearly 950,000 people from Rwanda.

While the model consistently outperformed previous demographic tools on test data, the authors acknowledge important limitations. The framework reveals that uncertainty remains exceptionally high in developing regions, particularly across sub-Saharan Africa. Because the underlying historical reporting in these areas is so sparse, the AI's confidence bounds explicitly highlight where on-the-ground data collection is most urgently needed by the international community.

By releasing their full dataset and trained models openly, the researchers have provided a transparent, reproducible foundation for the future of demographic research. Moving forward, this computational framework could pave the way for even finer geographical tracking, eventually allowing policymakers to monitor internal, climate-driven displacement on a highly localized scale.

Reference:
Gaskin, T., Abel, G.J. Deep learning four decades of human migration. Nature (2026). https://doi.org/10.1038/s41586-026-10611-7