topics|Cloudy with a chance of revolt

AI conflict prediction

After two decades of disappointing results, advances in machine learning and large language models have revived attempts to forecast political violence. Models are trained on past conflicts and fed indicators such as crime, public health, labour strikes, weather, economic data, political developments and social-media discourse. Increasingly, satellite, drone and surveillance-camera imagery is used to spot patterns in street life and protest behaviour.

Major systems

Integrated Strategic Forecasting (ISF): built by RAND, completed February 2026; put the chance of regime collapse or replacement in Iran by end-2026 at 20% in an Economist commissioned forecast.
CAST: developed by ACLED, uses indicators including infant mortality and peace-talk occurrences to predict bouts of organised political violence up to six months out. Used by UN agencies and the Dutch foreign ministry.
VIEWS (Violence & Impacts Early-Warning System): led by Håvard Hegre at the Peace Research Institute Oslo; adopted by the UNDP and the EU's diplomatic service.
IOM displacement model: published its first forecast on April 1st 2026, predicting that drought, floods and fighting would drive 304,362 people in Somalia from their homes over three months.
University of the German Armed Forces in Munich is building a "prediction hub" for Germany's defence ministry.

The best predictor of conflict is past conflict (ACLED's Katayoun Kishi). Models perform poorly at predicting the onset of new conflicts, partly because available data fail to capture true triggers (growing inequality may matter less than changing perceptions of it). Disinformation campaigns muddy social-media signal.