Towards Insider Summarization for Mediation Instead of Moderation: Examining Wikipedian Views on Key Elements of Discussion Summaries

In Wikipedia, conflict plays a vital role in refining knowledge. Yet some disputes persist without clear resolution or rule violations, resulting in long, circular discussions without moderator intervention. These deter neutral and third-party editors and often end only when one side gives up. In such cases, what is needed is not moderation but mediation-support that helps participants voice perspectives and reach consensus. However, mediation is hard to implement manually, as it requires summarizing complex discussions with a full grasp of both content and the Wikipedia community's culture. One potential solution is delegating this task to AI-but what content should be included in summaries that mediate and facilitate disagreements? Through a three-phase interview study with 14 Wikipedians, we examined how they read and interpret discussions, create their own summaries, and evaluate large language model (LLM)-generated summaries presented as technology probes. Our findings show that Wikipedians expect summaries to include key discussion elements-usernames, topics, arguments, sources, editor behavior, rule violations, and resolution-with particular emphasis on community-related context that reflects insider understanding. Based on these insights, we apply a new theoretical framework grounded in computer-mediated communication (CMC) theories such as common ground theory, warranting theory, and social presence theory to examine Wikipedians' interpretations of user identity. We also discuss design implications for discussion summaries on Wikipedia talk pages.