NetherlandsBellingcat, a digital journalism organization, has developed a machine learning model to help identify incidents of civilian harm on Telegram during the Russia-Ukraine war. Between February 2022 and September 2025, they collected over 2,500 verified cases of civilian harm, using these to train the model. The system significantly reduces the time needed to sift through vast volumes of user-generated content, allowing researchers to focus on verification rather than discovery. The model considers both direct harm (deaths, injuries) and indirect impacts like mental trauma, displacement, and infrastructure damage. To improve accuracy, the team created a dataset of 5,848 confirmed harmful posts and 48,545 non-harmful posts, ensuring the model reflects real-world proportions. The project highlights the potential of AI tools in open-source conflict research while emphasizing ethical considerations.
Bias read (Center): The article focuses on the technical development and application of AI in identifying civilian harm during conflicts, which is a politically sensitive issue due to its implications for humanitarian efforts and military accountability. However, the tone remains objective, detailing methodologies and倫
