Spiral of Silence: How Neutral Moderation Polarizes Content Creation

Ying Bao, Jessie Liu · 2025

Read Spiral of Silence: How Neutral Moderation Polarizes Content Creation by Ying Bao, Jessie Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-19680.txt.

Introduction

Ying Bao and Jessie Liu model how even neutral content moderation can polarize what people create online. Users both write and read posts, differ in ideology, and dislike content from the other side. Moderation in their model targets only toxicity, never viewpoint, which is the usual fairness aim.

Removing toxic posts gives clean posts more exposure, a benefit majority creators partly enjoy. Yet the same shift raises out-group hostility toward minority creators, who do not share in the benefit. Creation from the minority then dries up, echoing the spiral of silence. The result shows a fair-seeming rule can leave minority users facing a more one-sided feed.

Catalog record

CREATORYing Bao, Jessie Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-19680

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Spiral of Silence: How Neutral Moderation Polarizes Content Creation Thanks: The authors are listed alphabetically. We are grateful for the feedback received from participants at POMS, IIOC, the Marketing Science Conference, the CESifo Summer Institute on Digital Platforms, the Marketing Theory Seminar, and seminars at the University of Washington and University of Illinois at Urbana-Champaign. We also thank Rafael Jiménez-Durán, David Godes, Michael Keane, Peter Landry, Alexei Makarin, Cameron Martel, Sridhar Moorthy, Marcel Preuss, Mengze Shi, Shubhranshu Singh, Mateusz Stalinski, Michael Zhang, Yanhui Wu, and Pinar Yildirim for helpful conversations. All errors are our own. Ying Bao, Jessie Liu Abstract This paper investigates how content moderation affects content creation in an ideologically diverse online environments. We develop a model in which users act as both creators and consumers, differing in their ideological affiliation and propensity to produce toxic content. Affective polarization, i.e., users’ aversion to ideologically opposed content, interacts with moderation in unintended ways. We show that even ideologically neutral moderation that targets only toxicity can suppress non-toxic content creation, particularly from ideological minorities. Our analysis reveals a content-level externality : when toxic content is removed, non-toxic posts gain exposure. While creators from the ideological majority group sometimes benefit from this exposure, they do not internalize the negative spillovers, i.e., increased out-group animosity toward minority creators. This can discourage minority creation and polarize the content supply, ultimately leaving minority users in a more ideologically imbalanced environment: a mechanism reminiscent of the “spiral of silence.” Thus, our model offers an alternative perspective to a common debate: what appears as bias in moderation needs not reflect bias in rules, but can instead emerge endogenously as self-censorship in equilibrium. We also extend the model to explore how content personalization interacts with moderation policies. Keywords: Content Moderation, Polarization, Platform Governance, Toxicity 1 Introduction “The opinion of only part of the population seemed to be the opinion of all and everybody, and exactly for this reason seemed irresistible to those who were responsible for this deceptive appearance.” — Alexis de Tocqueville, L’Ancien Régime et la Révolution (1856, p. 259) Toxicity is a pervasive challenge on social media platforms, affecting millions of users daily. According to the 44 , 4 in 10 Americans have faced online harassment, and 71% support tighter platform rules on toxic content. This growing concern is reflected in emerging regulatory frameworks worldwide, including German Network Enforcement Act (NetzDG) and the European Union’s Digital Services Act (DSA). 1 1 1 https://digital-strategy.ec.europa.eu/en/library/code-conduct-countering-illegal-hate-speech-online Existing research on content moderation largely focuses on the demand side: how users respond when speech is restricted. Prior work has shown that content removal can trigger backlash ( 26 ) , increase engagement ( 28 ) , or shift user composition ( 36 ) . While these studies have advanced our understanding of moderation’s impact on content consumption, they often overlook a supply-side question: How the anticipation of moderation shapes what content is produced in the first place? Our study takes this supply-side perspective as its starting point. This focus is increasingly vital in today’s social media landscape. Moderation policies are no longer perceived as occasional or exceptional: they are institutionalized, expected, and often legally mandated ( 2 ) . As platforms implement preemptive tools such as automated flagging and removal, and as creators adapt to avoid penalties or backlash, anticipatory behavior becomes more relevant than reactionary responses alone. In this regime, it is crucial to understand not only which content gets removed, but also what content is never created at all. A key motivation for our approach comes from a puzzling empirical pattern: opposite moderation policies can sometimes produce similar outcomes. When Reddit banned several toxic communities during “The Great Ban” in 2020, some former toxic users became more active, but most, especially the less-toxic ones, posted less or stopped altogether ( 9 ; 10 ) . Conversely, when Twitter relaxed its rules in 2022, hate speech nearly doubled ( 21 ) . These two cases illustrate that moderation does more than remove harmful content: it changes the entire environment in which users create content. On social media, users are not merely passive consumers of content, they are also strategic creators who weigh the risks and rewards of content creation. Despite growing attention to polarization, most existing research conflate ideological division with toxicity, treating both as a single notion of “offensiveness.” This theoretical conflation obscures the distinct incentives and consequences that arise from ideology-based versus toxicity-based preferences. In practice, however, algorithm-based moderation systems mainly targets the latter. For instance, widely used Google’s Perspective API are designed to be ideologically neutral, trained on data across the political spectrum and calibrated to detect toxic language rather than viewpoint ( 48 ) . Yet, content that is non-toxic in language can still provoke severe backlash when it expresses ideological disagreement, depending on who is reading. For example, TikTok influencer Leo Skepi, with over 4 million followers, faced strong backlash after stating that brands should not be blamed for not carrying all sizes. 2 2 2 https://time.com/6965324/leo-skepi-tiktok-clothing-size Disentangling these two types of preferences is therefore conceptually important, not only to understand how users engage with content, but also how they decide what to create. Our paper formalizes this distinction along two aspects of user preferences: (i) a vertical preference capturing aversion to overtly toxic content, and (ii) a horizontal preference capturing aversion to ideologically opposed content. This framework allows us to identify the conditions under which ex ante “neutral” moderation, by reweighting anticipated engagement, can produce systematically non-neutral outcomes ex post. In this paper, we develop a model in which users act as both consumers and (potential) creators of content, differing along two dimensions: their ideological position and their propensity to produce toxic content. Content creation depends on both intrinsic motivation and external engagement, whether positive or negative, from ideologically aligned (in-group) and opposing (out-group) audiences. A central feature of our framework is explicitly modeling affective polarization : the extent to which users respond emotionally to the ideological identity of content sources, rather than to the content’s tone or language itself ( 24 ) . Affective polarization gives rise to systematic in-group favoritism and out-group animosity : toxic posts often receive some tolerance within ideological groups, whereas even civil content can trigger backlash simply because they come from the “other side.” Recent experimental studies also confirm this growing trend: users tend to evaluate otherwise identical content more negatively when it comes from ideological out-group members ( 55 ) . Alongside this, our model captures how ideological imbalance , i.e., unequal group sizes within the user base, creates asymmetry in who receives more validation versus hostility. Together, these two forces give rise to what we term as relational externality : any policy that alters exposure, such as moderation or personalization, reshapes the balance between in-group favoritism and out-group animosity that governs creators’ incentives. In other words, a policy aimed at regulating one dimension of content (e.g., toxicity) can indirectly influence other aspects (e.g., ideology) of content creation by altering the social (relational) environment in which content is rewarded or penalized. Our analysis offers two key insights. First, when both affective polarization and ideological imbalance are high, moderation amplifies in-group favoritism and out-group animosity through increasing

Read the complete text