In September 2013 Reuters published an article about the work of censors at Sina Weibo. The article talks about the work of a microblog censor, including the large number of posts that they are expected to review per hour, the low pay, and the poor ongoing career prospects. (The low pay and conditions might have something to do with the glut of competition: a Chinese government official has previously suggested that 1 in 10 Beijingers work in propaganda services.)
Reuters article puts something close to a human face on work that underwent intensive quantitative study by independent researcher Tao Zhu and his colleagues at Bowden College, Rice University, and the University of New Mexico. The team sought to understand the reactiveness of Weibo’s censorship regime. To do this they developed a protocol that allowed them to track the response of censors to posts by Sensitive Users.
Sensitive Users were those who had previously lodged posts containing key words that had previously been (temporarily) blocked. The researchers added to their initial list of 25 users by penciling in users who had been re-blogged five or more times by these users on the assumption that they are also likely to fall into the category of Sensitive Users; users who had more than five posts deleted within “a period of time” were included in the study.
The researchers final list contained 3,567 Sensitive Users whose accounts they would monitor. They opened 300 fake Weibo accounts and used Tor circuits to cycle through IP addresses so they could monitor the 50 most recent posts of their Sensitive Users at a rate of once per minute without their connections being blocked by Weibo’s servers. They then implemented a secondary protocol to test whether a post that disappeared from the timeline of a Sensitive User was deleted by the user or by Weibo. Posts deleted by users themselves returned a “post does not exist” error. Posts deleted by Weibo were tagged as “permission-denied deletions”: Weibo’s servers indicated that the post still existed but that it was not publicly accessible. This protocol allowed the researchers to detect instances of censorship within 1-2 minutes of a post being deleted.
The team found that 25% of censored posts were deleted within 5-30 minutes, and that 90% of deletions occurred within 24 hours. They estimated that Weibo would need 1,400 censors working simultaneously 24-hours a day in order to process the approximately 70,000 new posts lodged each minute. This led them to (not unreasonably) surmise that Weibo used a combination of automatic and manual review of posts to facilitate censorship.
The researchers identified nine techniques that Weibo appears to use for censoring posts:
1. Explicit filtering: posts containing banned phrases are instantly rejected, and the user is notified that the post violated “regulation rules”.
2. Implicit filtering: when a user lodges a post containing a potentially sensitive phrase they notified that there is a delay due to “server data synchronization”; the post is likely then checked manually.
3. Camouflaged posts: these are instances where a censored post appears for the user on their timeline as but (without their knowledge) cannot be viewed by other users.
4. Backwards reposts search: the researchers found instances where re-posts of a censored post were found and deleted en masse.
5. Backward keyword searches: censors appear to retrospectively search for posts containing phrases that have subsequently become sensitive; this lead to a noticeable spike in censor activity.
6. Monitoring specific users: it appears that the posts of specific users are monitored more closely for censorable content.
7. Account closures: Weibo sometimes closes the accounts of users who post sensitive content.
8. Search filtering: Weibo maintains a list of phrases that are excluded from its search function.
9. Public timeline filtering: posts on sensitive topics are filtered out of the public timeline.
Different problems, different solutions
The researchers say that the Weibo censorship system maintains multiple lists of sensitive phrases, which accounts for the different responses to the use of different sensitive phrases. For example, one reference list contains phrases that result in a post being automatically deleted; another list contains phrases that lead to a post being flagged a post as requiring manual review and either approval or deletion.
Combined with China’s new laws that provide for jail terms of up to three years for users that post socially-disruptive “rumours” online, and the censorship processes applied to long form blogs, the Party has in place a highly developed censorship system that helps to maintain the harmony of the Internet in China.