The documents presented on preventhate.org this time explore the issue of hate speech from three complementary perspectives: detection technology, contextual understanding, as well as intervention & response.

For improved efficiency, one paper introduces a LoRA-tuned BERTweet AI architecture to boost the accuracy of text-based hate speech detection models. Tackling a different media format, another document presents the MM-HSD system, which uses both visual and text data to more accurately detect hate speech in video content. Moving to a linguistic challenge, a third study compares deep learning models using IndoBERTweet to find optimal performance for Indonesian Twitter hate speech detection. On a macro level, another paper examines the potential of AI to both detect toxic speech and contribute to governing a healthier ‘digital commons.’

In terms of contextual challenges, research shows that the inclusion of humor complicates hate speech perception, emphasizing the critical role of social context and user response. Finally, regarding human response and intervention, one document argues that effective hate regulation requires collective, community-driven action that goes beyond simple platform policies and laws. Another study analyzes the psychological and behavioral coping mechanisms of individuals who witness versus those who receive digital hostility, providing a framework for support. The last document, on Meta Oversight explains how the Meta board is advancing a new, broader standard for hate-speech moderation: rather than focusing solely on imminent threats (like violence or discrimination), it assesses longer-term social harms even when there’s no clear incitement or direct intent.

https://preventhate.org/2025/11/17/new-on-preventhate-org-policyinstitute-net-17-november-2025/