Researchers at KInIT have released MultiSocial, a benchmark dataset built to test how well detection tools can spot AI-generated text in a social media environment where content is usually shorter, more casual, and linguistically diverse.
Unlike earlier studies that mostly focused on English and long-form content, MultiSocial addresses a research gap by covering 470,000+ posts in 22 languages from five platforms (X, Telegram, WhatsApp, Gab, and Discord). The dataset mixes human-written content with outputs from seven multilingual large language models (LLMs). The team evaluated 17 detection methods, finding that fine-tuned models adapt well to social media data and consistently outperform zero-shot approaches.
Learn more about the findings on Vigilant’s website here.
