Evaluating the Effectiveness of Deplatforming as a Moderation Strategy on Twitter
This blog post summarizes a paper on understanding the effects of deplatforming offensive influencers on Twitter that will be presented at the 24th ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW). This paper received a Best Paper Honorable Mention Award at CSCW.
CONTENT WARNING: This blog post contains some offensive keywords.
Over the past five years, many controversial influencers with thousands of followers on mainstream social media sites like Facebook, Twitter and YouTube have been banned. These bans have been referred to as ‘deplatforming’. Once deplatformed, these influencers are barred from making another account using their real name. The most popular instance of such bans occurred when social media sites deplatformed the former President Donald Trump. That incident triggered a host of important conversations about delimiting the regulation power of platforms and navigating tensions between libertarian tendencies and authoritarian practices in platform governance.
In this paper, we sidestep this ethical debate and instead focus on the long-term consequences of deplatforming as a moderation intervention. Using case studies of 3 far-right, extremist influencers — Alex Jones, Milo Yiannopoulos, and Owen Benjamin, we ask — what happens after offensive individuals with large following are deplatformed? Do their popularity decline? Do their ideas continue to flourish? How do their supporters change their behavior? By answering these questions, we provide empirical insights and a computational framework that, when combined with the ethical theories currently being developed, can help community managers make more informed decisions about whether and when to issue deplatforming.
To answer our research questions, we examined observational data from Twitter through a temporal analysis of (1) tweets directly referencing deplatformed influencers, (2) tweets referencing their offensive ideas, and (3) all tweets posted by their supporters. For each influencer, we limited our data collection and analyses to the tweets posted in the period six months before and after their deplatforming. Working with over 49M tweets, we chose metrics that include posting volume and content toxicity scores obtained via the Perspective API. To account for ongoing temporal trends, we then used interrupted time series (ITS) regression analyses and Wilcoxon signed-rank tests to examine variations in levels of activity and toxicity. The above figure summarizes the computational framework we used to conduct our analyses.
We determined that deplatforming disrupted discussions about influencers: posts referencing each influencer declined significantly, by 91.77% on average. Additionally, the number of unique users and new users tweeting about each influencer also diminished significantly, by 89.51% and 89.65% on average, respectively. Before this work, it was not clear whether or how deplatforming would affect the levels of posting activity about influencers. Certainly, these influencers continued to actively promote their views elsewhere. In fact, their deplatforming on Twitter was widely covered by multiple news outlets and provided them greater visibility, which could lead to a Streisand effect of drawing more attention to censored individuals. However, our results show that deplatforming significantly reduced the activity levels around these influencers. This suggests that deplatforming helped reduce the overall impact of these influencers on the platform.
We analyzed the spread of many offensive ideas popularized by deplatformed influencers. Our ITS analyses show that even after controlling for temporal trends, deplatforming helped reduce the spread of many of these anti-social ideas and conspiracy theories. This suggests that deplatforming diminishes not just the influence of banned individuals, but also of their ideas.
Our data also show that the deplatforming action significantly reduced the overall posting activity levels of supporters for each influencer: the median drop in the volume of tweets posted by supporters averaged 12.59%. Finally, deplatforming significantly reduced the overall toxicity levels of supporters of each influencer: across the three cases, the median decline in toxicity score of tweets posted by supporters averaged 5.84%. Since the influencers we studied were deplatformed at different times, it is unlikely that the changes we observed are reflective of isolated Twitter-wide trends.
We found that deplatforming increased the prevalence of some offensive ideas we tested in this study. Additionally, although deplatforming helped reduce the overall activity and toxicity levels of supporters, a small group of supporters significantly increased both their activity and toxicity levels. Thus, platforms should attend to how deplatforming may impact the activities of other users associated with the banned accounts and regulate their activities when necessary. Our data collection and methodological approach may prove helpful in finding relevant keywords to track as well as identifying supporters and analyzing their change in behaviors.
As our findings show, deplatforming influencers reduced the posting activity levels of hundreds of their supporters. Therefore, it might not be in the financial interests of platforms to conduct deplatforming. Many critics have raised concerns about the financial benefits from advertising dollars that are potentially tied to allowing toxic content to remain on these platforms. However, platforms should recognize that when they allow people who promote toxic speech to spread their views in the name of free speech, they are degrading women, minorities and other vulnerable groups and minimizing their dignity. Considering these concerns, it is vital that platforms clarify their commitment to respecting the dignity of all their users and deplatform offensive influencers when appropriate. As our analyses show, deplatforming can help reduce the spread of offensive ideas and diminish the toxicity in posts made by certain user groups. Therefore, judiciously using this strategy may allow platforms to address the problem of online radicalization, a worthy goal to pursue even if it leads to short-term loss in advertising dollars.
For more details about our methods, findings, and policy implications, please check out our full paper that will be published in Proceedings of the ACM on Human-Computer Interaction (CSCW) 2021. For questions and comments about the work, please drop an email to Shagun Jhaver at shagun.jhaver [at] rutgers [dot] edu.
Citation:
Shagun Jhaver, Christian Boylston, Diyi Yang, and Amy Bruckman. 2021. Evaluating the Effectiveness of Deplatforming as a Moderation Strategy on Twitter. Proc. ACM Hum.-Comput. Interact. 5, CSCW2, Article 381 (October 2021), 30 pages. https://doi.org/10.1145/3479525
