How to Use AI to Help Find Civilian Harm

Bellingcat describes a prototype machine-learning pipeline that ranked Telegram posts for likelihood of containing incidents of civilian harm in Ukraine: they assembled a labeled dataset (5,848 positive, 48,545 negative instances), engineered 893 numerical features including text embeddings and semantic similarity scores, evaluated several models (selecting XGBoost by PR-AUC), and discussed ethical considerations and operational deployment for aiding human verification.