Silenced on social media: the gatekeeping functions of shadowbans in the American Twitterverse

Abstract

Algorithms play a critical role in steering online attention on social media. Many have alleged that algorithms can perpetuate bias. This study audited shadowbanning, where a user or their content is temporarily hidden on Twitter. We repeatedly tested whether a stratified random sample of American Twitter accounts (n ≈ 25,000) had been subject to various forms of shadowbans. We then identified the type of user and tweet characteristics that predict a shadowban. In general, shadowbans are rare. We found that accounts with bot-like behavior were more likely to face shadowbans, while verified accounts were less likely to be shadowbanned. The replies by Twitter accounts that posted offensive tweets and tweets about politics (from both the left and the right) were more likely to be downtiered. The findings have implications for algorithmic accountability and the design of future audit studies of social media platforms.

Shaping opinions in social networks with shadow banning

“…Shadow banning policies are calculated daily, with the maximum mean shadow banning strength s network set to 0.05, and no limit to the shadow banning strength on each individual edge ( s edge = 1). The maximum mean shadow banning strength of 5% is chosen based on [ 2 ] that found 6.2% of sampled Twitter accounts were shadow banned at least once within a year of data collection. In addition, [ 56 ] estimated that between 0.5% to 2.3% of users were banned in the Twitter networks they studied.…”

“…The field study in [55] recruited 23,377 US-based adult Facebook users to assess the impact of modifying the polarity of content seen by users on their political polarization. The observational study in [2] audited a random sample of 25,000 Twitter accounts to identify if they were shadow banned. In addition to replicating the size of networks in these works, using a subgraph of our data also reduces the computational time of the simulations.…”

“…Shadow banning can serve as an effective strategy for maintaining the health of social media discourse. It was found that Twitter shadow banned accounts that exhibited automated or bot-like behavior, along with offensive posts [ 2 ]. While there are obvious benefits to shadow banning, it is not without potential drawbacks.…”

How Symbols Influence Social Media Discourse: An Embedding Regression Analysis of Trump’s Return to Twitter

“…The conflation of content’s lexical and symbolic properties also appears in research examining the consequences of platforms banning specific accounts (Rohlinger et al 2023). In this work, it is not always clear whether the ban has (or does not have) an effect on other users because it reduces users’ exposure to certain words or because a symbol—what an account represents—is removed (e.g., Jaidka et al 2023; Jhaver et al 2021). We see another example in recent work on social media content and polarization: Nyhan et al (2023) reduced Facebook users’ consumption of politically like-minded content by manipulating their feeds, but by doing so, they also decreased users’ encounters with like-minded friends and organizations, who were likely among their most familiar, trusted, and uplifting online interlocutors.…”

“…However, much of this research does not conceptualize nor operationalize content’s constitutive properties and their potentially distinct effects. For example, work examining social media platforms’ banning of user accounts (e.g., Jaidka, Mukerjee, and Lelkes 2023; Jhaver et al 2021; Rohlinger et al 2023) often does not specify whether the ban affects other users because it reduces those users’ exposure to certain words or because they no long encounter a specific symbolic meaning conveyed by the account.…”