Diyi Liu 刘迪一

ABSTRACT Social media platforms have become central actors in governing online speech during geopolitical conflicts, navigating complex governance demands from diverse stake-holding communities including states, civil society, and users. This paper explores these dynamics through an analysis of Meta, X (formerly Twitter), and TikTokduring the recent Russia–Ukraine war and Israel–Palestine conflict. We develop a framework categorizing platform strategies along two dimensions: activeness and alignment, leading to five response types: proactive, accommodation, avoidance, defiance, and disregard. Through analysis of 126 interaction events, we found three platforms’ distinct strategic approaches. Meta adopted the most proactive measures with clear inclinations to the Western states and calculated defiance against others. X’s strategy shifted dramatically after Musk’s takeover: from accommodation to defiance and disregard, though it made pragmatic accommodations under severe pressure. TikTok, risk-averse, and geopolitically vulnerable, frequently employs cautious and ambivalent avoidance and disregard strategies. The findings shed light on how corporate strategy, institutional background, and geopolitical positioning shape how platforms construct their legitimacy in content governance.

Content moderation is one of the earliest and most consequential large-scale applications of artificial intelligence, shaping both online safety and the boundaries of permissible speech. In practice, moderation operates as a sociotechnical human-AI system, combining automated flagging with human review, as automated systems remain too error-prone to operate independently. Leveraging newly mandated transparency disclosures under the European Union’s Digital Services Act (DSA), we conduct the first cross-platform audit of human content moderation workforce allocation across languages. Across six major platforms, we uncover substantial cross-lingual disparities in both language coverage and moderator staffing relative to the volume of user-generated content. While larger platforms such as YouTube and Meta employ moderators for many languages, millions of EU-based users on smaller platforms, including Twitter/X, post in languages without any dedicated human oversight. Among covered languages, staffing is often highly disproportionate to content volume, with English consistently prioritized over widely spoken Global Majority languages suchas Spanish, Portuguese, and Arabic. Where enforcement data permit analysis, we further document large cross-lingual disparities in individual moderator workload, with per-moderator daily decision volumes differing by more than an order of magnitude across languages. Together, these findings raise fundamental fairness concerns for both users and moderators, implying both an unequal user protection from online harms across linguistic communities, and an uneven distribution of the cognitive and emotional burdens of moderation labor. They further demonstrate that meaningful accountability for AI-assisted content moderation requires transparency not only about automated systems and enforcement outcomes, but about how human moderation capacity is allocated, justified, and sustained across languages.

To address the global challenge of online hate speech, prior research has developed detection models to flag such content on social media. However, due to systematic biases in evaluation datasets, the real-world effectiveness of these models remains unclear, particularly across geographies. We introduce HateDay, the first global hate speech dataset representative of social media settings, constructed from a random sample of all tweets posted on September 21, 2022 and covering eight languages and four English-speaking countries. Using HateDay, we uncover substantial variation in the prevalence and composition of hate speech across languages and regions. We show that evaluations on academic datasets greatly overestimate real-world detection performance, which we find is very low, especially for non-European languages. Our analysis identifies key drivers of this gap, including models’ difficulty to distinguish hate from offensive speech and a mismatch between the target groups emphasized in academic datasets and those most frequently targeted in real-world settings. We argue that poor model performance makes public models ill-suited for automatic hate speech moderation and find that high moderation rates are only achievable with substantial human oversight. Our results underscore the need to evaluate detection systems on data that reflects the complexity and diversity of real-world social media.

The governance of online speech is increasingly a battleground shaped by competing social expectations. This study investigates TikTok’s content moderation in Indonesia and Pakistan, two countries with vast market potential and delicate social and moral stances. Through document analysis and in-depth interviews with government officials, industry representatives, and civil society experts, it examines how stakeholders navigate normative and pragmatic considerations in global speech governance. The findings first highlight distinct regulatory approaches: Indonesia’s collaborative yet paternalistic model preferring fines over bans. It emphasizes administrative compliance through jurisdictional control over platform rules. In contrast, Pakistan’s defensive stance prioritizes infrastructure-level monitoring and restrictions, often resorting to platform bans to enforce control over moral and religious content. Unlike its Silicon Valley counterparts, TikTok demonstrates strategic compliance, deliberately sidestepping controversy by delegating sensitive decisions to state authorities and avoiding political roles. While normative consensus on appropriate content remains elusive, civil society organizations mediate crucial accountability relationships through strategic activism, coalition-building, and international networks. The study discusses the tensions and cost-benefit appraisals of each actor group, identifies essential principles for legitimate speech governance, and examines challenges in translating these principles into actionable frameworks.

Content moderation comes with trade-offs and moral dilemmas, particularly for transnational platforms governing borderline content where the boundaries of acceptability are subject to debate. While extensive research has explored the legality and legitimacy of platformised speech governance in established democracies, few address the complexities of less-than-democratic developing nations. Through sociolegal analysis and controversy mapping of TikTok’s localised moderation in South and Southeast Asia, the study examines how major actors negotiate the shifting boundaries of online speech. The analysis reveals that neither the platform nor regional states are well-positioned to achieve sound governance of borderline content. Primarily, TikTok localises its moderation based on pragmatic necessity rather than moral obligations, deliberately sidestepping contentious political controversies. Governments demonstrate determined efforts to control online discourse, leveraging legal uncertainty to further political agendas. Thus, local content governance often relies on vague rationales around securitisation and morality. The contradictory motives of (de)politicising borderline moderation seemingly counterbalance each other, yet in practice lead to an accountability vacuum where legitimate interests are sidelined. Given the lack of normative common ground, ensuring procedural justice and encouraging civic participation are essential to counteract rhetoric that rationalises imposition of certain speech norms hinging on imbalanced political power.