State & Geopolitical Censorship on Twitter (X): Detection & Impact Analysis of Withheld Content DOI: XXXXXXX.XXXXXXX Conference: The 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Ceox, Seoul, Korea ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered
Yusuf Mücahit Çetinkaya, Tuğrulcan Elmas · 2025
Read State & Geopolitical Censorship on Twitter (X): Detection & Impact Analysis of Withheld Content DOI: XXXXXXX.XXXXXXX Conference: The 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Ceox, Seoul, Korea ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered by Yusuf Mücahit Çetinkaya, Tuğrulcan Elmas (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-2508-13375.txt.
Introduction
Yusuf Mücahit Çetinkaya and Tuğrulcan Elmas measure what state-ordered withholding does to accounts on X, the site once called Twitter. Their paper from twenty twenty-five gives the first numeric study of this censorship form. Two patterns anchor the data: Russian accounts withheld in the European Union, and Turkish accounts blocked inside Turkey. Posting rates barely moved under the bans.
Engagement fell hard where the ban zone matched the account's home crowd. Likes and reposts dropped by a clear share, and follower growth nearly stalled. Russian accounts with outside audiences kept growing. The team also trains a classifier to flag accounts at risk of withholding. Tweet text carries nearly all the signal. Scores land in a solid range. The study feeds debates on platform rule, speech, and state pressure.
Catalog record
| CREATOR | Yusuf Mücahit Çetinkaya, Tuğrulcan Elmas |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-13375
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
State & Geopolitical Censorship on Twitter (X): Detection & Impact Analysis of Withheld Content DOI: XXXXXXX.XXXXXXX Conference: The 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Ceox, Seoul, Korea ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered computing Social media CCS: Social and professional topics Censorship CCS: Information systems Clustering and classification Yusuf Mücahit Çetinkaya, Tuğrulcan Elmas Abstract . State and geopolitical censorship on Twitter, now X, has been turning into a routine, raising concerns about the boundaries between criminal content and freedom of speech. One such censorship practice, withholding content in a particular state has renewed attention due to Elon Musk’s apparent willingness to comply with state demands. In this study, we present the first quantitative analysis of the impact of state censorship by withholding on social media using a dataset in which two prominent patterns emerged: Russian accounts censored in the EU for spreading state-sponsored narratives, and Turkish accounts blocked within Turkey for promoting militant propaganda. We find that censorship has little impact on posting frequency but significantly reduces likes and retweets by 25%, and follower growth by 90%—especially when the censored region aligns with the account’s primary audience. Meanwhile, some Russian accounts continue to experience growth as their audience is outside the withholding jurisdictions. We develop a user-level binary classifier with a transformer backbone and temporal aggregation strategies, aiming to predict whether an account is likely to be withheld. Through an ablation study, we find that tweet content is the primary signal in predicting censorship, while tweet metadata and profile features contribute marginally. Our best model achieves an F1 score of 0.73 and an AUC of 0.83. This work informs debates on platform governance, free speech, and digital repression. 1. Introduction State and geopolitical censorship on social media platforms like Twitter (X) has become a critical and growing concern. As governments increasingly leverage digital spaces to control narratives and suppress dissent, the line between legitimate content moderation and political suppression continues to blur. This complexity is further deepened by shifting platform policies—often influenced by political and geopolitical pressures—that determine how and when content is withheld. For instance, state-affiliated account networks spreading propaganda (troll) were suspended under the Civic Integrity project by Twitter, and the account data were shared until 2022. Twitter previously labeled state-affiliated accounts, such as the notable Russian news agency Russia Today, and soft-banned them by excluding them from recommendation algorithms and decreasing their reach. These practices are now discontinued under Elon Musk’s ownership ( Titcomb 2023 ) . Our focus in this paper is withholding content, a common censorship practice in which the content does not violate the platform’s terms of service, but the platform censors it by withholding it upon a legal request in the requesting jurisdiction, instead of suspending. How much Twitter accommodates or does not accommodate these requests is still controversial ( Lamensch 2024 ) . Elon Musk’s ownership has renewed public attention to the withholding content, as it contradicts his proclaimed vision of absolute free speech ( Rohlinger et al. 2023 ; Hickey et al. 2023 ; Robinson 2024 ) . This contradiction became most evident in March 2025 when X withheld İstanbul mayor Ekrem İmamoğlu’s account of 10 million followers and those of his supporters including university students, after his presidential candidacy was announced by the main opposition party in Turkey, intensifying debates over freedom of speech on X ( Guerin 2025 ) . Research on social media censorship has explored government takedown requests ( Tanash et al. 2015 ; Varol 2016 ; Voinea 2022 ; Caesmann et al. 2024 ; Gosztonyi et al. 2025 ) , and several studies have examined the role and effects of automated moderation in account suspension or shadow bans ( Volkova and Bell 2017 ; Chowdhury et al. 2021 ; Toraman et al. 2022 ; Pierri et al. 2023 ; Elmas 2023 ; Jaidka et al. 2023 ; Diskin 2024 ; Çetinkaya et al. 2024 ) . Yet, much of this work has relied on event-based or snapshot analyses, often isolating specific campaigns or focusing solely on individual tweets rather than users’ longitudinal behavior. To the best of our knowledge, our study is the first to quantify the sustained impact of state-imposed censorship on user activity and engagement over time. We use two datasets in our analysis. The first is a preexisting dataset containing censored content up to 2021, including takedowns in Germany, France, Russia, India, and Turkey. Among these, Turkey stands out due to the prevalence of accounts spreading militant propaganda ( Elmas et al. 2021 ) . The second is a new dataset we collected by focusing on the EU’s decision to censor Russian propaganda accounts over the Russia-Ukraine war ( Clayton 2022 ; Geissler et al. 2023 ) . These cases offer clear and high-impact examples of how geopolitical tensions shape social media governance. Our research questions are the following: RQ1: What is the impact of censorship on users’ activity and the engagements they receive? RQ2: Is censorship predictable based on content and profile signals? Our main contribution is the first systematic analysis of state-specific withholding on social media—capturing not just whether content is censored, but how users alter their posting behavior and engagement patterns when their tweets are selectively suppressed. We complement the analysis by a predictive task that models censorship at the user level by aggregating sequences of tweets, instead of treating each tweet in isolation. This approach provides a holistic view of how withholding unfolds over time, and it demonstrates, via an ablation study, that raw tweet content alone is the dominant predictive signal for withholding decisions, showing that metadata and profile features contribute little. 2. Data Censorships until 2021: The first dataset is collected using the Internet Archive’s Twitter Stream Grab, which contains 1% of all tweets between September 2011 and June 2020, from which all tweets with a non-empty "withheld in countries" field are extracted, yielding 583,437 censored tweets from 155,715 unique users ( Elmas et al. 2021 ) . Fully censored accounts are identified via the Twitter User Lookup API and an inference heuristic, resulting in 4,301 entirely withheld users. To support analysis, about 22 million non-censored tweets (including retweets) from users with at least one censored tweet are also collected. 98% of the censorships occur in Turkey, Germany, France, India, and Russia. 2022 EU Censorship of Russian Propaganda Accounts: The second dataset is collected using the 10% Decahose Twitter API between 2022 and 2023. We identified the accounts that were blocked in all EU countries and the UK through the “withheld in countries" field provided by the Twitter API, which are all Russian propaganda accounts as this was the only such instance where these countries unanimously censored accounts. We identified 38 accounts including Russia Today, Sputnik, and Ruptly. We collected all tweets by them, resulting in 148,037 tweets before and 123,685 tweets after withholding in March 2022. 3. RQ1: Impact Analysis Methodology: We analyze the impact of censorship on the accounts’ behavior and engagements by comparing the respective measures before and after the withholding event. We repeat this comparison for a control group. We employ the Wilcoxon signed-rank test to assess whether the changes before and after, for each measure are statistically significant.