Bots into the Fediverse

Social bots are a known problem in today’s society. They are influenced by a vari-ety of factors, ranging from the presence of bots to a lack of interaction betweenbots and users. This paper proposes a cross-platform approach for the detectionof social bots based on profile metadata and text embeddings, applied to Twit-ter, Mastodon, and Bluesky user accounts. The resulting model achieves 97.15%accuracy in a four-class classification task, outperforming several establishedbaselines, including graph-based and federated approaches while being compu-tationally efficient. The primary contribution of this work is the demonstrationthat user features can support effective bot classification across heterogeneousand decentralized environments, demonstrating the feasibility of cross-domaingeneralization at scale. We additionally present a novel dataset that combinesself-identified bot and non-bot accounts from decentralized platforms.