Bellingcat has developed a machine learning model to identify and prioritize Telegram posts that may contain evidence of civilian harm during conflicts, particularly in Ukraine since February 2022. The organization collected over 2,500 verified incidents of civilian harm and built a dataset using 5,848 confirmed posts containing harm and 48,545 posts without harm to train the model. The approach reduces the time needed to sift through vast amounts of social media content, allowing researchers to focus on verification rather than discovery. The model uses natural language processing techniques like multilingual transformers and cosine similarity to analyze text, while also considering metadata such as post timing and engagement. The project highlights both the potential of AI in humanitarian research and the challenges of organizing large volumes of user-generated content.
Bias read (Center): The article provides a technical explanation of a methodological innovation in open-source research on civilian harm, focusing on the development of an AI tool. It does not take a stance on political issues, nor does it favor one side in a debate. The framing remains neutral, emphasizing the utility
