Check out this new article by Yukun Yang, Robin Lange, Yidi Zhang, and Joseph B Walther on a simple but under tested question: how do other people’s reactions to online hate shape what hate posters do next. The paper aims to test two competing ideas. One says approval fuels more hate. The other says rejection and being ignored threatens belonging and can push people to escalate.
The team used real behavioural data from Gab, a platform known for hosting lots of hate content. They started with the Gab Scrape dataset (a subset of GabLeaks) covering posts from August 2016 to May 2018. From 22,790,465 total posts, they filtered out very short posts (fewer than 10 words), leaving 12,685,078 posts for scoring. They then measured toxicity on a 0 to 1 scale using Detoxify, treating scores above 0.5 as hateful. That produced 1,054,039 hateful original posts, from which they randomly sampled 1,000 posts written by 779 authors. For 985 of these posts, the authors had later activity, allowing the researchers to track 1,227,756 subsequent posts by the same users.
The key outcome was change in toxicity: first, the toxicity of the user’s next post compared with the sampled hate post; second, the average toxicity across the next three months compared with the sampled hate post. For responses, they counted Likes and Dislikes, plus written replies. Replies were manually coded in context to decide whether they affirmed the hate post or negated it, using a conversation based coding scheme. Reliability was checked with Krippendorff’s alpha, and the final dataset included 523 coded replies.
Here is the punchline. Likes and affirming replies did not increase later toxicity. When Likes and affirmations were combined as overall approval signals, more approval predicted a small drop in toxicity in the next post. No response at all also predicted a drop in toxicity in the next post. Dislikes were the clear risk factor: more Dislikes predicted higher toxicity in the next post and across the next three months. When Dislikes and negating replies were combined as disapproval signals, they also predicted increased later toxicity.
Policy wise, this is a warning about relying on public disapproval and some forms of counterspeech on hate friendly platforms. Visible rejection can backfire. Tools that reduce engagement and visibility, including moderation that limits feedback loops, may work better than piling on. For research, the paper shows why you need message sequences and conversation structure, not just cross sectional samples, and why platform context matters when testing theories about social reinforcement.