AI Facilitation of Online Disagreement

AI has considerable potential to facilitate online discussions and reduce conflict, but designers of such systems face the risk of algorithmic aversion. Past studies have found that people judge identical messages less favorably when they believe they were written by AI. However, as AI becomes more integrated into everyday life and people gain experience with AI facilitators, this aversion may diminish. It also remains unclear how much the content of a facilitator's message matters relative to its human or AI label, and whether the facilitation strategies people prefer actually improve the conversation. Finally, because most people observe rather than participate in online discussions, it is important to understand whether AI facilitators are perceived differently by observers and participants. To address these questions, we conducted two complementary studies examining perceptions of different facilitation strategies and their effects on actual conversations. In a vignette experiment (N = 388), we crossed the facilitator's label (AI or human) with ten facilitation strategies. In a live conversation experiment, an LLM used five of these strategies to facilitate 20-minute text chats between 188 pairs of people who disagreed on political and non-political topics. Participants preferred the kind of facilitator they had experienced. After seeing an AI facilitator, 30% wanted one in future discussions, against 7% after seeing a human facilitator. The AI label lowered ratings of the facilitator's warmth and competence, but the strategy mattered far more, explaining 11% of the variance in ratings against less than 1% explained by the label. Participants preferred facilitators that showed understanding over facilitators that reminded them of community norms. However, the strategies people preferred did not reliably change the conversations. Observers and participants in live conversations ranked the strategies similarly. In summary, the AI vs. human label matters far less than what the facilitator actually says. A well-designed process can improve user perceptions more than enough to offset algorithmic aversion.