This study explores an under-studied layer of Chinese Internet censorship: how Chinese Internet companies censor user–generated content, usually by deleting it or preventing its publication. Systematic testing of Chinese blog service providers reveals that domestic censorship is very decentralized with wide variation from company to company. Test results also showed that a great deal of politically sensitive material survives in the Chinese blogosphere, and that chances for its survival can likely be improved with knowledge and strategy. The study concludes that choices and actions by private individuals and companies can have a significant impact on the overall balance of freedom and control in the Chinese blogosphere.
By mid–2008, 253 million Chinese had gone online, surpassing the United States to become the country with the most Internet users (China Internet Network Information Center [CNNIC], 2008). By the end of 2007 there were 47 million bloggers in China with 72 million blogs, according to official statistics (CNNIC, 2007). By the end of 2008 Chinese Internet users were found to be spending more time online than Internet users in any other country. They were also found to be more likely to contribute to various kinds of online social networking sites — blogs, forums, chatrooms, photo or video–sharing Web sites, etc. — than people in all other countries surveyed except Korea and France (TNS Global Interactive, 2008). These statistics help to explain why the Internet has become the front–line battleground in China’s new “informational politics” (Yang, 2008).
In late 2007 Cai Mingzhao, a vice minister in charge of government information policy and regulation, emphasized that the Chinese online media of all forms must “have a firm grasp of correct guidance, creating a favorable online opinion environment for the building of a harmonious society” (Bandurski, 2007). Thousands of “Internet police,” deployed in many cities, are only one weapon used by the Chinese government in its battle the Chinese people’s hearts and minds (Xiao, 2007). Tens of thousands of overseas Web sites are blocked to users of domestic Chinese Internet services. Private citizens in every city and province are enlisted as volunteers or paid commentators to “guide” online conversations in a pro–government direction or to act as watchdogs, reporting anti–government conversations to the authorities (Bandurski, 2008b). Managers and employees of Internet companies — both domestic and foreign — are also expected to do their part in preventing China’s online discourse from getting out of hand (Pan, 2006a).
Chinese Internet censorship has received a great deal of attention by the international media, Western governments, international human rights activist community, and increasingly by the academy. Most of that attention, however, has focused on one part of China’s Internet censorship system: filtering. Internet filtering is the process by which users accessing the Internet from a particular network are blocked from visiting certain Web sites. Filtering can be done at various levels: the household; local business or residential networks; Internet service providers (ISPs); or, at the regional network or national gateway level. The same filtering techniques — and often the same software — are used by Western corporate offices, schools, and by a range of governments (Villeneuve, 2006). In China, filtering is achieved by plugging “blacklisted” Web site addresses and keywords into the routers and software systems controlling Internet traffic across Chinese domestic networks and at the gateway points through which information travels between the domestic Chinese Internet and the global Internet (Clayton, et al., 2006). China’s filtering system is widely known as the “Great Firewall of China” — a phrase coined informally by bloggers and Internet users. The international media often inaccurately conflates the “Great Firewall” with China’s official Golden Shield Project: a much broader project focused on surveillance, data mining and the upgrading of Internal public security networks, of which Internet filtering is only a very small part (August, 2007; Fallows, 2008; Walton, 2001; Klein, 2008). Filtering is the Chinese government’s primary method of blocking access to sensitive content hosted on overseas Web sites.
While a 2005 study by the Open Net Initiative found China’s Internet filtering system to be “the most sophisticated in the world” (OpenNet Initiative, 2005), this layer of censorship is imperfect because the “Great Firewall” can be “scaled” by proxy servers, secure tunneling, and other circumvention methods. As a result, the U.S. government and many non–governmental organizations hoping to promote free speech in China have invested substantial resources into the dissemination of information about circumvention tools to Chinese Internet users (Ha, 2006). Circumvention on its own, however, cannot solve the whole problem of access to information or open the door to a truly free public discourse in China. It does nothing to address the separate problem of domestic Internet censorship. For content on Web sites hosted on computer servers inside China, circumvention tools are irrelevant because content has been deleted or prevented from existing.
Lokman Tsui (2008) has suggested that the focus on the “Great Firewall” and Internet filtering by Western scholars, policymakers, media, and activists is due to their misguided tendency to view China through “Iron Curtain” Cold War–era paradigms. Others have pointed out that filtering is the one aspect of Chinese Internet censorship which foreigners visiting China are most likely to encounter personally, while censorship affecting the domestic Chinese–language Internet is much less apparent to visitors and outsiders. Furthermore, overseas–based free speech activists themselves — such as the Radio Free Asia, Human Rights Watch, Reporters Without Borders — have all had their own Web sites blocked by Chinese Internet filtering, making filtering the most immediate concern for them and their causes (MacKinnon 2008c). Whatever the reason for the focus on filtering and lack of attention to other kinds of censorship, this paper aims to help address the imbalance by shedding light on another part of China’s Internet censorship system: The process of domestic Web site censorship by which domestically hosted content is deleted completely or prevented from being published in the first place. The whole process is carried out almost entirely by employees of Internet companies, not by “Internet police” or other government officials. This study focuses specifically on one small piece of this domestic censorship system: How blog service providers (BSPs) censor blogs written by their Chinese users.
Blogging gained critical mass in China in 2004 and 2005. Because all the major international BSPs (Blogspot, Typepad, Wordpress.com, etc.) were blocked by the “Great Firewall,” most Chinese bloggers — who like most bloggers worldwide are not technically skilled enough to arrange their own Web–hosting and install their own blogging software — were forced to publish on domestic Chinese BSPs. By 2006, concerned about the role played by blogs, chatrooms and forums in the 2005 anti–Japanese street protests, authorities had created a system of regulations and obligatory “self discipline” pledges in hopes of compelling Web companies to keep user–generated content from going beyond certain limits (MacKinnon 2008a).
All companies running Web sites in China — portals, search engines, social networking services, chatrooms, forums, blogs, video– or photo–sharing Web sites, etc. — are now required to comply with government censorship demands in order to keep their business licenses (Reporters Without Borders, 2007). Politically sensitive content is deleted from the Web by company employees, or by computer programs written by company employees, either in response to official directives or often simply in anticipation of trouble (Weiquan Wang [Chinese Human Rights Defenders], 2008; Reporters Without Borders, 2007). In 2007 one Chinese blogger even tried to sue his BSP for censoring content, which he argued was not illegal – to no avail. The suit was thrown out of court (Olesen, 2007; Dickie, 2007).
Web sites hosted outside of China, containing Chinese–language content targeted at a mainland Chinese audience, are asked to prevent the publication of certain politically sensitive content, or face the possibility of being blocked. While some foreign companies have opted not to comply and thus forego Chinese market opportunities, others have complied to varying degrees. Google, Yahoo! and Microsoft all offer censored versions of their search engines to the Chinese market in order to maintain good government relations and their business operations in China (Human Rights Watch, 2006; MacKinnon, 2008b).
Recent studies of domestic Chinese search engines and foreign–branded search engines serving the Chinese market (specifically Baidu, Yahoo! China, MSN, and Google.cn) conducted by Human Rights Watch and the University of Toronto’s Nart Villeneuve reveal a great deal of variation from company to company in both the extent and the methods of censorship (Human Rights Watch, 2006; Villeneuve, 2008). These studies confirm anecdotal reports that while government regulatory bodies issue directives to companies about what kinds of content should be controlled, the finer details of implementation are left to the companies (Reporters Without Borders, 2007). The regulatory goal does not appear to be 100 percent deletion of all information or opinions portraying various parts of the government and Chinese Communist Party (CCP) in a negative light — a goal that is in any case unachievable on the Internet (MacKinnon, 2008a). Rather, rewards and punishments are meted out based on the extent to which Internet companies successfully prevent groundswells of public conversation around politically inflammatory topics that might inspire a critical mass of people to challenge Communist Party authority (Weiquan Wang, 2008; Reporters Without Borders, 2007).
From time to time, “keyword lists” maintained by companies — lists of sensitive words or phrases that depending on the circumstances may trigger removal of just the words themselves, censorship of entire postings containing those words, or as part of an alert system for human editors tasked with moderating large amounts of content — are leaked to the public (Xiao, 2004; Pan, 2006b; Weiquan Wang, 2008). These lists are not given directly to the companies by government regulators; rather they are developed by the companies themselves and sometimes shared within the industry, in reaction to on–going advisories, warnings, and complaints from regulators (Reporters Without Borders, 2007; Weiquan Wang, 2008). In spite of this censorship, recent quantitative research by Chinese media scholar Ashley Esraey, comparing material published in the Chinese blogosphere to Chinese newspaper content, reveals that the blogosphere is a much more freewheeling space than the mainstream media.
“Compared to the content of mainstream, traditional media, blogs are much more likely to contain opposing perspectives and criticism of the state.”
Also notable “was the absence of regime propaganda in the blogosphere.”  In order to evade the BSP’s internal censors, Chinese bloggers frequently deploy satire, euphemisms, literary allusions, vague or coded phrases, and even graphics to convey critical messages (Esarey and Xiao, 2008). Chinese bloggers who write about current events and public affairs are also inclined to be realistic and pragmatic when deciding what to write, making decisions about what kinds of postings would bring too much trouble and aren’t worth the risk (Pan, 2006a).
As of this writing, no systematic testing of BSP censorship appears to have been conducted in China or anywhere else. This paper is a first attempt to fill in that blank space. By testing fifteen different BSPs serving the Chinese market, our research team sought to answer three questions: How great is the variation in the quantity of material censored by different Chinese BSPs? To what extent are conversations about politically sensitive subjects able to survive in the Chinese blogosphere in spite of censorship? How — and to what extent — do censorship methods vary from company to company?
Our tests yielded some interesting answers: First, censorship levels across 15 different BSPs varied even more than expected. Second, a great deal of politically sensitive material survives in the Chinese blogosphere, and chances for survival can likely be improved with knowledge and strategy. Third, censorship methods vary greatly from company to company, implying that companies do have at least some ability to make strategic choices. These choices are not only about how to balance relationships with government and users, but also about the extent to which BSPs value user rights and interests.
Since no similar testing of blog censorship appears to have been done in China or anywhere else, our research team invented a new methodology through trial and error. Hopefully our methods will be refined and improved upon as more research is done on how user–generated content is censored by Internet companies.
The testing of Internet filtering at the Internet service provider (ISP) and network level — such as the filtering tests conducted by the OpenNet Initiative in China and around the world — is a largely automated process relying heavily on network interrogation by special software applications. Results are analyzed with the help of computer programs. While the initial list of URLs and keywords for testing must be selected by Chinese speakers with up–to–date knowledge of China’s political and social situation, the actual testing can be carried out by computers and the analysis can be done largely by automated processes and non–Chinese speakers (Open Net Initiative, 2005). Testing of censorship by blog hosting services, however, is a different matter. This kind of testing can only be done by people with a high level of Chinese reading comprehension, in addition to a good understanding of China’s contemporary socio–political developments. Our team found no way to automate the process, making it very time–consuming.
After roughly two months of experimentation, refinement, and discussion, our team devised a time–consuming methodology that we followed over an eight–month testing period, from early February 2008 till late September 2008. Tests were conducted intensively in February–April and July–September . The first testing period included the mid–March Tibet demonstrations, crackdown, and aftermath. The second testing period covered the run–up and aftermath of the Olympics as well as the beginning of the “tainted milk scandal” in which a vast amount of Chinese milk products were found tainted with the chemical melamine.
Anonymous author accounts were created on 15 different commercial BSPs catering to mainland Chinese bloggers. In alphabetical order they are: Baidu, Blogbus, BlogCN, iFeng, Mop, MSN Live, MySpace, Netease, QZone, Sina, Sohu, Tianya, Tom, Yahoo! China, YCool. Passages of one–three paragraphs from a variety of sources were posted across all 15 BSPs.
After consultation with Chinese journalists, bloggers, and media experts, we designated 50 subject categories that merited testing (See Table 1 for a breakdown of those categories). Throughout the testing period, articles or excerpts of articles were selected related to these 50 subject categories that had been published online in a wide variety of sources: blogs, forums, overseas dissident Web sites, mainstream Chinese news sites (such as sina.com, sohu.com, and xinhuanet.com — the Web site of the official government mouthpiece, the Xinhua News Agency), and overseas Chinese-language news sites (such as the Financial Times Chinese site (http://www.ftchinese.com/), the BBC Chinese Web site (http://www.bbc.co.uk/chinese/), Reuters Chinese (http://cn.reuters.com/), and the Wall Street Journal Chinese (http://chinese.wsj.com/). While we tested one article excerpt about the banned religious group Falun Gong and one article excerpt related to the 1989 Tiananmen Square crackdown in order to observe how the different Web services handle the most sensitive subjects, these subjects were not the focus of this study. Chinese bloggers inside China generally expect those subjects to be censored and reprisals for discussing these topics openly online are also widely feared. While this reality is a serious affront to Chinese bloggers’ freedom of speech, it is already well known that discussion of these subjects on the mainland Chinese internet is not possible (Sabbagh, 2006). The aim of this study was to focus on subjects bloggers inside China are more likely to be trying to write about, and thus to get a better sense of where the boundaries lie and how far they can be pushed.
After selecting a piece of content for testing, I posted it on an internal password–protected Web site specifically set up for censorship test management. The excerpt or “content unit” selected for testing was given a unique number and posted into the Web site, along with the URL of the original article, followed by the full text of that article in case it failed to remain at the original Web site. Members of our testing team would then follow a set procedure for each content unit, testing that content on each of the 15 BSPs in the following manner:
- Log into the blog of one BSP, copy and paste the content unit into the “back end” edit window of the blog; take a screenshot;
- Hit “publish.” If the content was blocked from publication, or “held for moderation” take a screenshot of what kind of error message or other message appeared;
- If the content was not blocked from publication, take a screenshot of what the blog post looks like when the author is logged in to the system;
- Log out and check whether the content is still visible to the public, not just the logged–in author. Take a screenshot;
- Check back 24–48 hours later to see if the blog post is still visible. Take a screenshot showing either the still visible post, or the error message saying “this page does not exist,” or whatever else can be seen;
- Access the the blog post on a mainland Chinese ISP to see whether it is accessible (i.e., not filtered) on at least one Chinese domestic ISP . Take a screenshot. (This step was primarily a way to check whether any of the services were geo–filtering their material as will be discussed further in the “Findings” section.)
- Upload all of these screenshots into a database according to the unique number assigned to the content unit, along with descriptive comments noting any interesting or unusual circumstances surrounding the situations in which censorship occurred.
This testing process was highly labor–intensive and time–consuming; one initial test “round,” not including the follow–up checks, would usually take an experienced tester approximately three–four hours. Test rounds that could not be completed in under a week were thrown out as invalid.
Overall 124 tests were conducted, with 16 thrown out as invalid due to tester error or inconclusive results, leaving 108 valid. At least one item was tested per subject category, with greater emphasis placed on subjects related to current events. Some items were tagged with more than one category since we were testing excerpts of articles and blog posts whose subjects often didn’t fall neatly into a single category. Articles or excerpts of articles were used rather than simple sentences or keywords because censorship conducted by at least some BSPs is not an entirely automated process, involving manual checking by a staff member to determine the context in which certain sensitive phrases or keywords are used. Greatest emphasis was placed on “sudden incidents” — a Chinese euphemism for breaking news stories of a sensitive nature such as a demonstration, riot, act of violence, or manmade disaster — because bloggers who are interested in public affairs tend to be most interested in current events and breaking news, and breaking news (along with the spikes of conversation that develop around some breaking news topics) is also of greatest interest to government regulators. Altogether, 23 of the test items were tagged as “sudden incidents.” Ten tests related to Tibet in some way and 16 tests related to the Olympics. Corruption–related subjects were also emphasized, with 15 tests.
Finding 1: The extent of censorship by each BSP varies drastically.
A certain amount of variation in censorship levels from BSP to BSP was expected, based on casual observations and reports of bloggers’ experiences with censorship that inspired this study in the first place. We did not, however, expect the drastic variation that we found.
As can be seen in Figure 1, out of 108 valid tests conducted across 15 blog hosting services, the most vigilant company censored 60 blog posts (56 percent of the total). The second most vigorous company censored 44 (41 percent), and the third most vigorous censored 34 (32 percent). At the other end of the scale, the least vigorous blog host censored only one piece of content, the second most liberal censored only three, and the third most liberal censored nine.
Figure 1: Instances of censorship on 15 different Blog Service Providers (BSPs)
(company names omitted due to concern for government repercussions).
The names of the BSPs have not been published alongside these aggregate results. People who work in the Chinese Internet sector have expressed strong concern that published results of an independent academic study showing who censors more than whom could be used as a tool for reward and retribution by regulating authorities. The purpose of this study is to increase global understanding of how Chinese BSP censorship works. Naming company names alongside aggregate censorship results would certainly enable deeper examination of why certain companies censor a great deal more than others. However, this benefit is outweighed by the costs, not only to individuals working in Chinese Internet companies, but also to companies that are clearly trying to respect their users’ interests and rights over regulatory demands to the extent possible while not losing their business licenses.
Even without naming companies along with their censorship rankings, the results are revealing. The wide variation in levels of censorship confirms that censorship of Chinese user–generated content is highly decentralized, and that implementation is left to the Web companies themselves. Thus it has so far been possible for at least some Chinese Web companies (even some large, popular brands) who run user–generated content platforms to remain in business despite inconsistent and patchy levels of censorship.
Finding 2: Some politically sensitive material can survive in the Chinese blogosphere.
A substantial amount of politically critical content survives in the Chinese blogosphere. Not a single content item was censored by all 15 BSPs. Even the item about Falun Gong was censored by only 13 BSPs — the highest number of BSPs to censor the same item. An item on Tibet independence was censored by 12 BSPs. An excerpt of the Dalai Lama’s open letter to the Chinese people and an item about the “Tiananmen mothers” Web site were both censored by nine BSPs. An item about a riot in Weng’an county, Guizhou province was censored by seven BSPs. A commentary mocking China’s legislature, calling it an entertainment show, was censored by four BSPs. An item on the growing gap between rich and poor in the Chinese countryside was censored by two BSPs.
Table 1 breaks down how many content units tested were censored by how many BSPs. Of 108 valid tests, 21 content items were censored by none of the 15 BSPs. Twenty–four content items were censored by only one BSP .
Table 1: Tests broken down by number of censored results (15 BSPs tested). Same item censored by 0 of 15 BSPs 21 tests Same item censored by 1 of 15 BSP 24 tests Same item censored by 2 of 15 BSPs 15 tests Same item censored by 3 of 15 BSPs 6 tests Same item censored by 4 of 15 BSPs 13 tests Same item censored by 5 of 15 BSPs 4 tests Same item censored by 6 of 15 BSPs 8 tests Same item censored by 7 of 15 BSPs 6 tests Same item censored by 8 of 15 BSPs 3 tests Same item censored by 9 of 15 BSPs 3 tests Same item censored by 10 of 15 BSPs 3 tests Same item censored by 11 of 15 BSPs 0 test Same item censored by 12 of 15 BSPs 1 test Same item censored by 13 of 15 BSPs 1 test
Automated keyword–based censorship systems sometimes over–compensated, causing a Xinhua News Agency article about President Hu Jintao’s inspection tour of coal–fired power plants to be censored by four BSPs .
As mentioned in my earlier description of our testing methodology, content items — eventually posted on 15 BSPs — were catalogued into 50 subject categories (see Table 2).
Table 2: 50 subject categories. 1. Sudden incidents 11. Foreign trade/invest 21. Macau politics 31. local leaders 41. natural disaster 2. overseas political events 12. finance and economy 22. AIDS 32. religion 42. economic measures 3. Olympics 13. problems inside govt. ministries 23. health 33. minorities 43. dissidents (not jailed) 4. historical issues 14. corruption 24. crime 34. political reform 44. censorship/surveillance 5. leftist critiques 15. relocation 25. city govt. policies 35. “rights defense” 45. opposition parties 6. military/security 16. environment 26. provincial govt. policies 36. critiques of govt. human rights policy 46. National People’s Congress 7. foreign policy 17. Three Gorges Dam 27. natl. govt. policies 37. political arrests 47. labor issues 8. anti–Japanese 18. HK politics 28. media/tech policy 38. independence movements 48. migrant workers 9. anti–U.S. 19. Taiwan politics 29. national leaders 39. regime change 49. economic disparity 10. N. Korean refugees 20. Taiwan independence 30. provincial or city leaders 40. Falun Gong 50. Tibet
Items tagged with “crime,” “foreign trade and investment,” and “anti–Japanese” were not censored by any BSP. Posts related to Falun Gong, the 1989 Tiananmen crackdown, and Tibet independence were heavily censored, as expected. Items labeled with the following subject categories also received heavy censorship (more than 10 instances in which a post with one of these labels was censored): sudden incidents, Olympics, Tibet, corruption, dissidents, ethnic minorities, national leaders, political arrests, labor issues, censorship/surveillance, critiques of government human rights practices, independence movements, involuntary relocation, “rights defending” cases, historical issues, and misc. religious matters. Contrary to expectation, content items tagged “Taiwan politics” and even “Taiwan independence” were lightly censored — perhaps due to the fact that Taiwan elected a more moderate president in the spring of 2008 as well as the fact that pro–Taiwan independence politicians were plagued by corruption scandals, with both developments being reported openly in the mainland Chinese media.
In our effort to discover where the censorship fault lines lie, testing emphasized topics related to controversial current events and breaking news. Of the 23 tests of content related to “sudden incidents,” subjects included: developments in the poisoned milk powder case, local protests or clashes with police (including Tibet but also in other places), explosions in Xinjiang, the knifing of tourists in Beijing during the Olympics, natural or man–made disasters. Two BSPs censored 14 of 23 posts tagged with “sudden incidents,” and most censored at least five posts, although one BSP censored only one and one BSP censored none at all. Censorship was very inconsistent in terms of which BSPs censored what content. For example: four items related to the melamine poisoned milk powder scandal were tested. Two of the items were censored by two BSPs, one was censored by three BSPs, and one was censored by four BSPs. One BSP (iFeng) censored all four items; one BSP (Tianya) censored two of the four items (including a report which originated from the Web site of China’s national state–run television, about a visit by Premier Wen Jiabao to a store selling milk powder); Sina, MSN, Blogbus, Mop and Myspace censored one each.
Below are four more examples of just how inconsistently certain hot topics were handled across different BSPs:
Tibet: Ten tests contained Tibet–related content. Thirteen of the 15 BSPs censored at least two Tibet–related items. The most vigorous BSP censored eight out of 10 posts. Two BSPs censoring no Tibet–related items tested (See Figure 3 below).
Olympics: Sixteen tests contained Olympics–related content (including discussion of a Chinese female gymnast’s age, pressures felt by athlete Liu Qiang, lip–synching at the opening ceremonies, arrest of old ladies who applied to protest, etc.). The most vigorous BSP censored eight, while four censored none, and three censored only one of the 16 posts (See Figure 2 below).
Corruption: Fifteen tests contained material about corruption at various levels of government. The most vigorous BSP censored nine entries; one BSP censored none; all others censored between one–four posts (See Figure 2 below).
Dissidents: Five tests contained material about political dissidents. All services censored at least one post; the most vigorous censored four posts (See Figure 2 below).
Figure 2: Censorship of different kinds of content.
Some content directly opposing Communist Party rule was published successfully on all fifteen BSPs. One example was an excerpt of a hard–hitting essay by Bao Tong, former secretary to the late Zhao Ziyang who was deposed in 1989, in which he argued that the Olympics were being used to bolster the one–party state they would not bring China any closer to democracy as many foreigners had hoped. While Bao’s name was removed from the excerpt for testing purposes (to see whether the language itself rather than the author’s notoriety would trigger censorship), it contained references to democracy and the Universal Declaration of Human Rights, and criticized the single–party regime throughout. Another example was an excerpt essay by the Guangzhou–based blogger Yang Hengjun discussing why a multi–party political system with regular democratic elections is preferable to China’s current political system. None of the 15 BSPs tested censored either of those two posts.
Findng 3: There is wide variation in censorship methods.
Censorship methods vary substantially. What’s more, sometimes companies use different methods depending on the specific nature of the content. Some companies also choose to be more transparent or honest than others when communicating with users whose blog posts have been censored. The basic censorship methods can be categorized as follows:
1. Tester is prevented from posting at all — Upon clicking “publish,” the tester is presented with an error message of some form, with varying degrees of explanation but usually implying that the content was sensitive in some way (See Figure 3). Details are never given, providing an explanation as to what exactly the offending content was or why it was un–publishable. Industry sources have confirmed that posts censored in this way are blocked via an automated system triggered by keywords, phrases, or even whole passages that are plugged into the system by administrators. This method was used at least once by 11 of 15 BSPs: Baidu, BlogCN, Mop, iFeng, Myspace, Netease, QZone, Sina, Tianya, Tom, and Yahoo China.
Figure 3: Screenshot of Baidu error message:
“Sorry, publication of your article has failed. Your article contains inappropriate content, please check.”
The article is about a clash between police and residents of a small city in Yunnan province, taken from the official Xinhua News Agency.
2. Post is “held for moderation” — Upon clicking “publish,” the tester was presented with a message indicating that the content is being held for approval, apparently the result of an automated process triggered by the use of keywords. This often happened on the same services that have also prevented publication of other posts, indicating that some services categorize different types of content at different sensitivity levels, to be handled differently. In some cases the content held for moderation was eventually published, indicating that a human being reviewed it and determined that the content was acceptable. In other cases the content was “held for moderation” indefinitely (See Figures 4a and b). This method was used at least once by 10 of 15 BSPs: Baidu, BlogCN, iFeng, Mop, Netease, QZone, Sina, Tianya, Tom, and Yahoo! China.
Figure 4a: Screenshot of iFeng notice (on article advocating Maoism):
“Your blog article has been submitted, it needs approval before appearing, thank you.”
Figure 4b: Screenshot of iFeng front page private view (when author is logged in).
Headline appears with a note alongside it: “Sorry, this blog article is undergoing approval..”
Headline and note don’t appear in public view (when author is logged out). Post remains in moderation forever; never appears in public view.
3. Post is published in “private view,” but is never visible to the public. The tester is able to publish the post, and it is visible on the blog’s front page but only to the author when the author is logged in to the system. The post cannot be seen when the author logs out of the system and it is not publicly visible. This technique was used less frequently, at least once by three of the 15 BSPs: Mop, Netease, and Tom.
4. Post is successfully published at first, but deleted or “unpublished” some time later — usually within approximately 24 hours, although over weekends it could sometimes take as long as two days before a blog post would be taken down (See Figures 5a and b). Industry sources have confirmed that in these cases the content is flagged by the internal software system due to the presence of keywords; it is then reviewed by someone who then decides whether to remove, or un–publish, the post in question. This method was utilized at least once by 10 of the 15 BSPs: Blogbus, BlogCN, Mop, MySpace, QZone, Sohu, Sina, Tianya, Yahoo! China, and YCool.
Figure 5a: Screenshot of blog post on Sina.com.
Article excerpt about explosion in Xinjiang province is successfully published in public view.
Figure 5b: 24 hours later, this error message appears at the same URL.
5. Sensitive keywords or phrases are replaced with “***” but the post is otherwise published — This method of censorship, an automated process triggered by keywords, was observed on only two of the 15 blog hosting services: Blogbus (its main form of censorship in most though not all cases), and Yahoo China (which utilized this method in addition to several of the others above). For example: on Blogbus, the name of President Hu Jintao and mentions of the Tibetan independence movement were replaced by asterisks, even when the text we tested was copied from Xinhua News Agency propaganda articles (See Figure 6). While replacement of sensitive keywords with asterisks is Blogbus’ primary means of censorship, it does occasionally remove entire posts after publication.
Figure 6: Blogbus replaces some of the characters from the phrase “Tibet independence” with “***”.
Article is excerpted from BBC Chinese, about how some foreign Tibet independence protesters were kicked out of China.
6. The content is successfully published, but blocked to viewers attempting to read it from inside mainland China. Only Windows Live, the blog hosting service run by Microsoft, utilized this method (See Figure 7a and b). This method is only possible when the content itself is hosted on servers outside China.
Figure 7a: Excerpt of an article about the poisoned milk powder scandal, publicly visible on Hong Kong ISP.
Figure 7b: “Connection close” message appears when attempting to access the same URL from a Mainland Chinese ISP.
Microsoft’s approach to geographically targeted censorship without actually removing content on Windows Live is the result of controversy over the company’s deletion of a popular Chinese blog in late 2005. After facing substantial international criticism by human rights groups, the press, and the U.S. Congress, Microsoft agreed with critics that outright deletion of content — rendering it inaccessible to global audiences, not just Chinese audiences — was not an acceptable practice for a global blog hosting platform. On the other hand, Microsoft executives in China had been informed that the service would be blocked on Chinese ISPs if they took a zero–censorship approach to politically sensitive content. The result was a compromise: objectionable content would be blocked to visitors using mainland Chinese Internet connections, but would remain published and visible everywhere else (MacKinnon, 2008a).
Some BSPs used primarily one method, but most services utilized a combination of different methods: blocking at publication stage, holding for moderation, or removal after publication. This indicates that many BSPs have set up complex systems in which different sets of keywords fall into different categories of sensitivity.
Different services also chose to explain — or not explain — censorship to their users in different ways. Nearly all BSPs censored an item about the “Tiananmen mothers,” (one of the few items censored by most but not all services), but different methods were used. Many but not all services blocked the content from publication, using pop–up messages similar to the one which appeared when a tester attempted to publish an article excerpt about the “Tiananmen mothers” on a MySpace blog: “Sorry, your post contains inappropriate content. Please delete the inappropriate content and repost, thanks.” QZone, on the other hand, allowed the post to be published in “private view” (visible only to the author when logged in) but the post is not publicly visible: in its place appears a message: “This message is being previewed, which will take 3 working days. Once approved it will be possible to view normally.” The post never appears.
The blogging platforms of Sina.com and Sohu.com, both major Beijing–based Internet companies, appeared to employ a combination of internal keyword flagging followed by human decision–making most of the time. Politically sensitive posts that were successfully published at first vanished from the blog within 24 hours. Interestingly, each company handled this de–publication differently. Sina simply removed the post without explanation. Attempts to access the post’s original URL produced an error message saying: “Sorry the blog address you have visited does not exist.” Sohu, on the other hand, un–pubished and demoted the post to draft status, enabling the original content to be retrieved by the author through the back–end author interface even though it could no longer be publicly seen. Posts that were un–published in this manner were also flagged with an explanatory message saying: “This blog post has been hidden.” A box immediately below it states: “Dear Sohu Blog Friend: Hello! We’re very sorry to inform you that this blog post, because of certain reasons, is not suitable for publication, and has been locked. You can access the original text and photos from this page. Thank you for your understanding and support of Sohu. Sohu Customer Service is available to you twenty–four hours per day at: [provides phone number and e–mail address].” At the blog post’s original URL, visitors are met with an error message that says: “This blog post has been hidden.”
These examples of how Myspace, QZone, Sina and Sohu handle censorship differently demonstrate that Chinese blog hosting companies can and do make choices about how honestly they communicate with users about censorship. Sohu’s honesty was noted by blogger–lawyer Liu Xiaoyuan, who in 2007 had tried unsuccessfully to sue Sohu for censoring content that Liu believed did not violate the terms of service. In December 2008 Liu wrote:
“In the past I sued Sohu for deleting my blog posts, but now I want to praise them. Sohu is the only BSP [blog service provider] that posts notices on my blog saying, ‘this post has been hidden/removed for certain reasons.’ As a result, when web users visit my Sohu blog, they can know that a post has been hidden by Sohu. I think Sohu is brave to do this. I also run blogs on Sina, ifeng, and others, but they simply delete blog posts without notifying my readers.” (MacKinnon, 2008d)
Thus, while Chinese companies are not able to defy government censorship demands altogether, test results show that they have a meaningful amount of leeway: not only in terms of how they respond to these demands, but also how they communicate with users about censorship of their works. It is unclear how much impact these have on user decisions about which blog hosting platform to patronize. Since these censorship differences are not advertised, it’s not clear whether many users are aware of these substantial differences. It’s also unknown what percentage of Chinese bloggers have had at least one blog post censored. Given that Chinese Internet users consider many factors when deciding which Web service to use (such as what their friends are using, attractiveness and user–friendliness of the user interface, other useful services and entertainment provided, etc.) it is difficult to speculate on whether Chinese Internet users would gravitate in large numbers to BSPs who have a reputation of lighter censorship and greater transparency. Further research needs to be done.
Based on follow–up interviews with people in the Chinese Internet industry who spoke on condition of anonymity, there seem to be a number of reasons for the wide variation in censorship practices. Some companies appear to be able to “get away” with less censorship than others. According to interviewees, this is due to a number of factors: It depends where the company is registered, because (as discussed in Section 3) a Web company is generally answerable to regulators at the city or provincial level in the city or province where the company is registered. It depends on how large and successful a company is, how much public attention its front–page portal receives, and how much competitive pressure the company is under. It also depends on who owns the company, who its investors are, and how much political pressure they are feeling in relationship not only to the Web business concerned but also other technology or media businesses they may have at play in the Chinese market. Finally, interviewees claimed that the amount of censorship carried out by a given BSP — and the way in which it is implemented — depended on the character, background, interests, and courage of individual editors hired to manage a given Web company’s blog portal. In several cases interviewees cited editors’ and managers’ journalism background as a factor: when asked why company X censored its blogs significantly less than company Y, one theory given was that the blog portal editor for company X was a journalist while the blog portal editor for company Y had a background in technology and business. All of these explanations, however, are only hypotheses, requiring more systematic research in order to be proven conclusively.
What is interesting about these explanations is that they paint a picture of domestic Chinese Internet censorship that in many ways mirrors the way that China’s traditional print media is controlled and regulated. It is well documented by scholars of the Chinese news media that some Chinese newspapers and magazines manage to “get away” with much more hard–hitting investigative stories than others. The reasons depend on various factors such as: where the head office of the publication or TV station is located; who owns the media property in question; who is the editor–in–chief, how well connected he or she is with powerful people in the government, what are his or her personal values and agendas, etc. (Bandurski 2008a; Hassid, 2008) In other words, despite the authoritarian nature of the Chinese state, people working in both old and new media are not entirely powerless: Individual choices, values, and actions by individual journalists are proven to have an impact on the quality and range of information and ideas available to the Chinese people. It would appear that in China’s new media, individual values and choices similarly help to shape the extent to which Chinese netizens are able to engage in an informed public discourse. On the other side of the coin, private and non–government actors collaborate with the government to muzzle each other for various reasons of self–interest, expediency, and even patriotism.
The findings of this study point to the need for more study — both of Chinese domestic Internet censorship as well as censorship in other countries. This study was highly experimental and limited in its scope, timeframe, and resources. Larger–scale testing would help to shed greater light on the way in which different kinds of content are censored. This study was limited to blogs; it did not test other forms of social networking Web sites, chatrooms or bulletin board systems (BBS), which are extremely popular and influential on the Chinese Internet, instant messaging, and mobile services. Surveys of Internet company employees would help to shed better light on the reasons behind the wide variation of censorship practices. Surveys of Chinese bloggers and Internet users would help us to answer questions such as: How often do average Chinese bloggers encounter censorship? What do they think about it? How have they reacted and in what ways have they modified their online behavior after encountering censorship? How do censorship practices impact his or her loyalty towards a particular BSP?
There are also some global research questions: Where else in the world is this kind of political censorship by web service companies of user–generated content happening? Companies in the West already censor for child porn, copyright violations and sometimes hate speech, but to what extent are Web companies in other countries besides China systematically complying with government demands to delete politically sensitive material? Will the “Censorship 2.0” model — in which governments demand censorship by Web companies — spread globally? Given how difficult it is to carry out such censorship consistently and effectively, and how much staff time and resources must nonetheless be taken up in attempting to implement it, would it be in companies’ commercial interests to resist or reject government efforts to delegate censorship to companies? Further research is needed in order to understand the global trends and emerging practices.
Improved knowledge of China’s domestic censorship system also helps to inform the work of activists working for greater freedom of speech, not just in China but also around the world. When a substantial amount of content is being deleted from the World Wide Web, promoting circumvention — while important — does not fully address the problem of Internet censorship. Bloggers and Internet users would benefit from more systematic information about strategies for successfully disseminating politically sensitive information via domestic Web sites on which content is frequently censored. Finally, free expression advocates should consider how an “Internet user rights” movement might push for greater transparency and accountability by Internet companies. The Vietnamese government has already announced that it will roll out similar Internet industry regulations to China’s . Global Internet companies have received censorship requests from governments around the world, including democracies such as Thailand, Turkey, and India. Consumers will need to put greater pressure on their service providers to resist government attempts to delegate political control and manipulation to the private sector.
About the author
Rebecca MacKinnon is an Assistant Professor at the Journalism and Media Studies Centre, University of Hong Kong, where she teaches online journalism. Her research and writing focuses on issues of online free speech, censorship, and citizen media, with an emphasis on China. She is co–founder of Global Voices, an international citizen media network, and a founding member of the Global Network Initiative. She is also a former CNN bureau chief in Beijing and Tokyo.
Email: rmack [at] hku [dot] hk
This study was made possible by a seed research grant from the University of Hong Kong. I am also grateful for the support and encouragement of Prof. Yuen–Ying Chan, Director of the Journalism and Media Studies Centre.
Research assistants John G. Kennedy and Shun–yi Lai spent countless hours conducting censorship tests. John G. Kennedy also provided useful feedback and many insightful ideas. Ben Cheng custom–built our test–management and database system, helped to manage the testing work, and contributed numerous ideas on the testing methodology. Without the hard work and dedication of these three people this project could not have been completed.
1. Esarey, 2008, p. 10.
2. The break was caused by the primary investigator’s unavailability; ideally testing would have been continuous throughout the February–September.
3. This was done either via mainland Chinese proxies or Chinese Tor exit nodes, or by people physically in mainland China when possible and practicable. See Villeneuve (2008) for more details about use of Tor nodes as a method of testing Chinese ISP behavior.
4. The distinction of being sole censor of one particular piece of content was shared by eight different BSPs in different tests.
5. Leaders’ names are often included in automated censorship systems, especially in cases where a company does not have adequate staff for manual vetting of content before deletion.
6. “Vietnam to regulate blogging” (2 December 2008), at http://chao-vietnam.blogspot.com/2008/12/vietnam-to-regulate-blogging.html, accessed 26 January 2009.
Oliver August, 2007. “The Great Firewall: China’s Misguided — and Futile — Attempt to Control What Happens Online,” Wired (23 October), at http://www.wired.com/politics/security/magazine/15-11/ff_chinafirewall, accessed 4 January 2009.s
David Bandurski, 2008a. “Garden of Falsehood,” Index on Censorship, volume 37, issue 2 (May), pp. 45–54.http://dx.doi.org/10.1080/03064220802081571
David Bandurski, 2008b. “China’s Guerrilla War for the Web,” Far Eastern Economic Review (July), at http://www.feer.com/essays/2008/august/chinas-guerrilla-war-for-the-web, accessed 4 January 2009.
David Bandurski, 2007. “State Council vice–minister reiterates control as top priority of Internet development in China,” China Media Project (4 December), at http://cmp.hku.hk/2007/12/04/763/, accessed 4 January 2009.
China Internet Network Information Center (CNNIC), 2008. 22th Statistical Survey Report on the Internet Development in China (July), abridged English version at http://www.cnnic.cn/en/index/0O/02/index.htm; full Chinese language version at http://www.cnnic.cn/index/0E/00/11/index.htm, accessed 4 January 2009.
China Internet Network Information Center (CNNIC), 2007. 2007 Nian Zhongguo Boke Shichang Diaocha Baogao [2007 Report on the Chinese blog market] (December); in Chinese only at http://www.cnnic.cn/html/Dir/2007/12/26/4948.htm, accessed 4 January 2009.
R. Clayton, S. Murdoch, and R. Watson, 2006. “Ignoring the Great Firewall of China,” In: G. Danezis and P. Golle (editors). Privacy Enhancing Technologies 2006, LNCS 4258 (December) pp. 20–35.
Mure Dickie, 2007. “China traps online dissent,” Financial Times (12 November), at http://www.ft.com/cms/s/0/ef0e7d64-9138-11dc-9590-0000779fd2ac.html, accessed 4 January 2009.
Ashley Esarey, 2008. “Political Discourse in the Chinese Blogosphere: A Quantitative Analysis,” paper presented at the 6th Annual Chinese Internet Research Conference (13–14 June, University of Hong Kong, Hong Kong).
Ashley Esarey and Qiang Xiao, 2008. “Below the Radar: Political Expression in the Chinese Blogosphere,” Asian Survey, volume 48, number 5 (September/October), pp. 752–772.http://dx.doi.org/10.1525/AS.2008.48.5.752
James Fallows, 2008. “The Connection Has Been Reset,” Atlantic (March), at http://www.theatlantic.com/doc/200803/chinese-firewall, accessed 4 January 2009.
K.O. Ha, 2006. “Piercing China’s Firewall: Hackers, Activists, Challenge Beijing’s Internet Police,” San Jose Mercury News (2 July).
Jonathan Hassid, 2008. “Controlling the Chinese Media: An Uncertain Business,” Asian Survey, volume 48, number 3 (May/June), pp. 414–430.http://dx.doi.org/10.1525/as.2008.48.3.414
Human Rights Watch, 2006. Race To the Bottom: Corporate Complicity in Chinese Internet Censorship (August), at http://www.hrw.org/reports/2006/china0806/, retrieved 4 January 2009.
Naomi Klein, 2008. “China’s All–Seeing Eye,” Rolling Stone (28 May), at http://www.rollingstone.com/politics/story/20797485/chinas_allseeing_eye, retrieved 4 January 2009.
Rebecca MacKinnon, 2008a. “Flatter World and Thicker Walls? Blogs, Censorship and Civic Discourse in China,” Public Choice, volume 134, numbers 1–2 (January), pp. 31–46.
Rebecca MacKinnon, 2008b “Asia’s Fight For Web Rights,” Far Eastern Economic Review (April), at http://feer.com/essays/2008/april/asias–fight–for–web–rights, accessed 4 January 2009.
Rebecca MacKinnon, 2008c. “The Chinese Censorship Foreigners Don’t See,” Wall Street Journal Asia (14 August), at http://online.wsj.com/article/SB121865176983837575.html, accessed 4 January 2009.
Rebecca MacKinnon, 2008d. “Studying Chinese Blog Censorship,” RConversation (November), at http://rconversation.blogs.com/rconversation/2008/11/studying–chines.html, accessed 4 January, 2009; Original Chinese blog post by Liu Xiaoyuan at http://blog.sina.com.cn/s/blog_49daf0ea0100bmb0.html, accessed 4 January 2009.
Alexa Olesen, 2007. “China tightens Internet controls,” Associated Press article appearing on MSNBC.com (14 October), at http://www.msnbc.msn.com/id/21268635/, accessed 4 January 2009.
OpenNet Initiative, 2005. Internet Filtering in China in 2004–2005: A Country Study, at http://www.opennetinitiative.net/studies/china/ONI_China_Country_Study.pdf, accessed 1 January 2009.
Phillip Pan, 2006a. “Bloggers Who Pursue Change Confront Fear And Mistrust,” Washington Post (21 February), at http://www.washingtonpost.com/wp–dyn/content/article/2006/02/20/AR2006022001304.html, accessed 4 January 2009.
Phillip Pan, 2006b. “Keywords Used to Filter Web Content,” Washington Post (18 February), at http://www.washingtonpost.com/wp–dyn/content/article/2006/02/18/AR2006021800554.html, accessed 4 January 2009.
Reporters Without Borders, 2007. “Journey to the Heart of Internet Censorship,” (October), at http://www.rsf.org/article.php3?id_article=23924, accessed 4 January 2009.
Dan Sabbagh, 2006. “No Tibet or Tiananmen on Google’s Chinese site,” Times Online (25 January), at http://business.timesonline.co.uk/tol/business/markets/china/article719192.ece, accessed 4 January 2009.
TNS Global Interactive, 2008. Key Insight Report 5: Digital World, Digital Life, Taylor Nelson Sofres, London (December), at http://www.tnsglobal.com/news/key-insight-reports/, accessed 4 January 2009.
Lokman Tsui, 2008. “The Great Firewall as Iron Curtain 2.0: The implications of China’s Internet most dominant metaphor for U.S. foreign policy,” paper presented at the 6th annual Chinese Internet Research Conference (13–14 June, University of Hong Kong, Hong Kong), at http://jmsc.hku.hk/blogs/circ/files/2008/06/tsui_lokman.pdf, accessed 4 January 2009.
Nart Villeneuve, 2008. “Search Monitor Project: Toward a Measure of Transparency,” Citizen Lab Occasional Paper, number 1, University of Toronto (June). at http://www.citizenlab.org/papers/searchmonitor.pdf, accessed 1 January 2009.
Nart Villeneuve, 2006. “The filtering matrix: Integrated mechanisms of information control and the demarcation of borders in cyberspace,” First Monday, volume 11, number 1 (January), at http://firstmonday.org/htbin/cgiwrap/bin/ojs/index.php/fm/article/view/1307/1227, accessed 4 January 2009.
Greg Walton, 2001. China’s Golden Shield. Montreal: International Centre for Human Rights and Development, at http://www.ichrdd.ca, accessed 1 January 2009.
Weiquan Wang (Chinese Human Rights Defenders), 2008. Zhongguo wangluo jiankong yu fanjiankong niandu baogao (2007) [Annual report on Chinese internet surveillance and censorship, and actions against censorship and surveillance] (10 July) at http://crd–net.org/Article/Class1/200807/20080710165332_9340.html, accessed 1 January 2009.
Qiang Xiao, 2007. “China censors Internet users with site bans, cartoon cop spies,” San Francisco Chronicle (23 September), at http://www.sfgate.com/cgi-bin/article.cgi?f=/c/a/2007/09/23/INCLS80LO.DTL, accessed 4 January 2009.
Qiang Xiao, 2004. “The words you never see in Chinese cyberspace,” China Digital Times (30 August), at http://chinadigitaltimes.net/2004/08/the_words_you_n.php, accessed 4 January 2009.
Guobin Yang, 2008. “Activists beyond Virtual Borders: Internet–Mediated Networks and Informational Politics in China,” First Monday, Special Issue number 7: Command Lines: The Emergence of Governance in Global Cyberspace, at http://firstmonday.org/htbin/cgiwrap/bin/ojs/index.php/fm/article/view/1609/1524, accessed 4 January 2009.
Paper received 17 January 2009; accepted 25 January 2009.
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
China’s Censorship 2.0: How companies censor bloggers
by Rebecca MacKinnon
First Monday, Volume 14, Number 2 - 2 February 2009
A Great Cities Initiative of the University of Illinois at Chicago University Library.
© First Monday, 1995-2014.