Echoes of Norms: Investigating Counterspeech Bots' Influence on Bystanders in Online Communities

TL;DR

A counterspeech strategy framework is developed and Civilbot is built for a mixed-method within-subjects study, highlighting when to intervene and how to do so through reasoning-driven and context-aware strategies.

Abstract

Counterspeech offers a non-repressive approach to moderate hate speech in online communities. Research has examined how counterspeech chatbots restrain hate speakers and support targets, but their impact on bystanders remains unclear. Therefore, we developed a counterspeech strategy framework and built \textit{Civilbot} for a mixed-method within-subjects study. Bystanders generally viewed Civilbot as credible and normative, though its shallow reasoning limited persuasiveness. Its behavioural effects were subtle: when performing well, it could guide participation or act as a stand-in; when performing poorly, it could discourage bystanders or motivate them to step in. Strategy proved critical: cognitive strategies that appeal to reason, especially when paired with a positive tone, were relatively effective, while mismatch of contexts and strategies could weaken impact. Based on these findings, we offer design insights for mobilizing bystanders and shaping online discourse, highlighting when to intervene and how to do so through reasoning-driven and context-aware strategies.

Figures (6)

• Figure 1: Sample interface of the simulated discussion platform, showing: (a) an excerpted question; (b) neutral answers; (c) a hate-speech post. After participants complete the pre-test questionnaire, the interface displays (d) a counterspeech message. Participants may "Like" the counterspeech or post their own comments in response to the hate-speech post. After finishing all interactions, they click "Home", complete the post-test questionnaire, and return to the homepage.
• Figure 2: The overall experiment procedure, including four phases: (A) Pre-survey, (B) Introduction, (C) Experiment sessions, (D) Post-survey.
• Figure 3: Overview of results for RQ1–R3. RQ1 shows overall effects on bystanders, Civilbot's roles for them, and perceived community-level mechanisms; RQ2 presents strategy-level effects and key interaction patterns; Design implications illustrate design insights of Civilbot from the perspective of bystanders, organized by when to counterspeak and how to counterspeak.
• Figure 4: Heatmap of the correlation between mean scores of different variables and the eight strategy groups. On the y-axis, Q/NQ indicates question or non-question sentence type, P/N indicates positive or negative tone, and C/A indicates cognitive or affective strategic intent. Variables on the x-axis: Q1 (convincing reason), Q2 (strong reason), Q3 (credibility), Q4 (importance), Q5 (overall agreement), Q6 (confidence in countering) and Q7 (willingness to participant)
• Figure 5: Heatmap of pairwise paired t-tests between counterspeech types across the three questionnaire measures. Significant comparisons are outlined with bold borders. Cell values represent effect sizes (Cohen’s d). Perceived quality = mean(Q1–2); subjective acceptance = mean(Q3–5); behavioral tendencies = mean($\Delta$Q6–7) ($\Delta$ = post $-$ pre).