Every Rose Has Its Thorn: Censorship and Surveillance on Social Video Platforms in China Jeffrey Knockel1,2, Masashi Crete-Nishihata2, Jason Q. Ng2, Adam Senft2, and Jedidiah R. Crandall1 1Dept. of Computer Science, University of New Mexico 2Citizen Lab, Munk School of Global Affairs, University of Toronto Abstract which suggests there may be general directives provided by the government on what content to censor, but the Social media companies operating in China face a com- companies have a degree of flexibility for how they im- plex array of regulations and are liable for content posted plement it [16, 25, 31]. Further exploring these obser- to their platforms. Through reverse engineering we pro- vations is challenging due to methodological constraints. vide a view into how keyword censorship operates on four The majority of studies on Chinese social media rely on popular social video platforms in China: YY, 9158, Sina sample testing in which researchers develop a set of con- Show, and GuaGua. We also find keyword surveillance tent suspected to be blocked by a platform, send the sam- capabilities on YY. Our findings show inconsistencies in ple to the platform, and record the results. This approach the implementation of censorship and the keyword lists introduces inherent bias as results are only as accurate as used to trigger censorship events between the platforms the overlap between the sample and the actual content fil- we analyzed. We reveal a range of targeted content in- tered. Another approach is to observe changes to a system cluding criticism of the government and collective action. (such as content deletion) over time. Studies using this These results develop a deeper understanding of Chinese method are typically limited to snapshots within a spe- social media via comparative analysis across platforms, cific period, which constrains longitudinal analysis. Ap- and provide evidence that there is no monolithic set of plications that implement censorship and surveillance on rules that govern how information controls are imple- the client-side (i.e., by the application itself rather than on mented in China. a remote server) present a unique research opportunity. Reverse engineering applications can reveal keyword lists 1 Introduction used to trigger censorship and surveillance. These lists are unbiased samples that provide comprehensive visi- The Chinese Internet is a complex ecosystem that in- bility into technical implementation and target content. cludes multiple layers of technical and regulatory infor- This study provides a broad look into keyword surveil- mation controls that are affected by actors with different lance and censorship across social video platforms positions of influence and responsibility. Extensive work (SVPs), a popular class of applications in China. SVPs has been done on China’s national-level Internet filtering combine real-time video streaming and social network- system, but this system is only one of many layers of in- ing features that enable users to broadcast content and formation control in the country. create interactive groups. One of the most popular uses Gaining a wider understanding of censorship and is broadcasting karaoke performances. SVPs are primar- surveillance in China requires analysis of its Internet plat- ily monetized through the sale of virtual goods (such as form developers and companies. These companies op- virtual roses) that users give to performers during broad- erate in a highly constrained regulatory environment in casts. While musical performances account for the ma- which they are responsible for the content on their ser- jority of revenues, SVPs are expanding to gaming, edu- vices and subject to fines and loss of operating licenses cation, financial analysis, and online dating applications. if they are found in violation. This governance model ef- Through reverse engineering, we identify client-side fectively pushes responsibility for information control to keyword censorship in four of the most popular SVPs: the private sector. YY, 9158, Sina Show, and GuaGua. In the case of YY Previous work has shown inconsistencies in how dif- we also find keyword surveillance capabilities. Our anal- ferent Chinese Internet companies conduct censorship, ysis reveals a dataset of 17,547 unique keywords, which 1 Company Product Reg. Users MAUs of Sina Show. Tian Ge reports user numbers as aggre- YY Inc. YY 861.4 mn. 117.4 mn. gates across its platforms and in 2014 had 245 million 9158 registered uses and 14.4 million MAUs [3, 4]. In July Tian Ge 245.0 mn. 14.4 mn. Sina Show 2014, Tian Ge went public on the Hong Kong Stock Ex- Jinhua not change. GuaGua 70 mn. Changfeng available Jinhua Changfeng Information Technology Co., Ltd., is a privately held company based in Zhejiang Province, Table 1: Social Video Users by Platform China that provides the GuaGua platform, which as of 2013 had 70 million registered users [9]. See Table 1 for a breakdown of user bases across SVPs. to our knowledge is the largest unbiased collection of cen- sorship keywords currently available. 1 Our main findings 2.1 Legal and Regulatory Environment in are as follows: China Inconsistencies in targeted content and implemen- The Communist Party of China (CPC) attempts to bal- tation between platforms: Comparing our SVP dataset ance the growth of information and communication tech- to previously collected chat client censorship keyword nologies in the country and limits on speech that can lists [16] allows for the first comparison of unbiased key- threaten its power [26]. In 2010 China’s State Council word samples across different industry segments. Our Information Office published what is considered the first analysis reveals limited list overlap between companies, government issued policy document on the Internet. It which substantiates previous findings that suggest com- includes a list of prohibited topics: panies are only given general directives from authorities and have a degree of flexibility in the implementation. endangering state security, divulging state Range of targeted content including criticism of the secrets, subverting state power and jeopar- government and collective action: While there is lim- dizing national unification; damaging state ited direct overlap in unique keywords, across lists we see honor and interests; instigating ethnic hatred or trends in the topics that are targeted including social is- discrimination and jeopardizing ethnic unity; sues, criticism of the government, and collective action. jeopardizing state religious policy, propagating These findings serve as a counterpoint to previous work heretical or superstitious ideas; spreading ru- from King et al. [21, 22] who posit that content related mors, disrupting social order and stability; dis- to collective action is heavily censored on Chinese social seminating obscenity, pornography, gambling, media while content critical of the government is often violence, brutality and terror or abetting crime; allowed to persist. humiliating or slandering others, trespassing Our findings provide strong evidence that there isno on the lawful rights and interests of others; and monolithic set of rules governing how information con- other contents forbidden by laws and adminis- trols are implemented in China and that developing holis- trative regulations [8]. tic understandings of the Chinese Internet requires com- Companies are held liable for content on their plat- parative analysis across platforms. forms and are expected to invest in staff and technol- ogy for ensuring compliance with government regula- 2 Background tions. Failure to comply with regulations can lead to fines or revocation of operating licenses. Complicat- The most popular SVPs in China include YY, Sina Show, ing matters is the vague language used in Chinese reg- 9158, and GuaGua. YY is developed by YY Inc. based ulatory documents. Terms like “disrupting social order in Guangzhou, China, and is the largest platform in terms and stability” are not clearly defined and punishments of user population. As of December 2014, YY had are meted out seemingly arbitrarily, thus pushing com- 861.4 million registered users and 117.4 million average panies and users to both over-censor and self-censor—a monthly active users (MAUs) [6]. In November 2012, phenomenon coined by Perry Link as “the anaconda in YY Inc. announced an initial public offering on the Nas- the chandelier” [24]. daq stock market. It is currently the only Chinese SVP In public filings for YY and Tian Ge both companies company to be traded on the US stock market. underline the risk that their businesses face from poten- Tian Ge Interactive Holdings Limited based in tial legal sanctions being brought against them for hosting Hangzhou, China owns and operates two SVPs: 9158 and prohibited content [2, 6]. The companies also highlight Sina Show. In 2010, Sina Corporation invested 10 mil- the risk of being affected by government campaigns such lion dollars (representing a 25% stake) in Tian Ge and as “Clean the Web 2014,” which was an government ef- provided the company a sole license for the operation of fort to crack down on the creation and dissemination of 2 pornographic content online. During this campaign, Sina then sent to the content monitoring team for further re- Corporation received notices regarding prohibited con- view. The company also generally describes audio mon- tent on its platforms and was subsequently fined 5.1 mil- itoring and keyword filtering systems. In addition, it pro- lion RMB, had