From Moderation to Mediation: Can LLMs Serve as Mediators in Online Flame Wars? (arxiv.org)

Recent research investigates the potential for large language models (LLMs) to act as mediators in online conflicts, transitioning from traditional moderation roles to more nuanced interpersonal engagement in digital discussions. This study introduces a dual-subtask framework for mediation, consisting of judgment—evaluating fairness and emotional dynamics—and steering, where LLMs generate empathetic responses to facilitate resolution. Utilizing a comprehensive dataset from Reddit, the authors implemented a multi-stage evaluation process that includes principle-based scoring, user simulation, and human comparison. The findings reveal that API-based models significantly outperform open-source counterparts in mediation tasks, demonstrating enhanced capability in reasoning and intervention alignment. This shift towards using LLMs as mediators underscores the growing interest in harnessing artificial intelligence for social good, particularly in promoting constructive dialogue and reducing online hostility. However, the study also notes the limitations of current LLM technologies, signaling a need for ongoing research to refine these models for effective emotional intelligence and empathy in diverse online interactions.