Abstract

Conversational moderation, intervening in conversations to encourage constructive behavior, is an effective alternative to banning users or deleting comments, which can exacerbate polarization by driving users toward echo chambers. However, it is challenging to scale, as human moderators are scarce and it is emotionally taxing to repeatedly respond to toxicity. In this paper, encouraged by the enhancement to conversational AI through developments in large language models, we study the potential for scaling conversational moderation through automatic moderation suggestions from moderator bots. To study the effectiveness of these suggestions independent of human mediation, we ask human evaluators to continue controversial conversations collected from Reddit with various moderator bots. We find that prompted large language models can provide specific and fair feedback to toxic behavior, but struggle …

Date

2023

Authors

HYUNDONG CHO, SHUAI LIU, DARPAN JAIN, BASEM RIZK, YUYANG HUANG, ZIXUN LU, NUAN WEN, JONATHAN GRATCH, EMILIO FERRERA, JONATHAN MAY