Tweeting Under Pressure

Wilson, Chen, ZhangWhen Fan Jiyue, party sec­re­tary of Lushan County in China, vis­ited the site of an earth­quake last April that claimed the lives of nearly 200 of his con­stituents, it was his wrist that drew the most atten­tion. A dis­tinct tan line sug­gested he’d removed a watch before being pho­tographed, which social media users quickly took as evi­dence of Fan’s cor­rup­tion because the watch was believed to cost well above the cus­tomary salary range for a country official.

The next day, Fan’s name was blocked from search results on Sina Weibo, the Chi­nese ver­sion of Twitter, and tweets con­taining the name were taken down within min­utes. Cen­sor­ship isn’t uncommon on the microblog­ging plat­form, which counts some 500 mil­lion users. Research sug­gests that up to 16 per­cent of all “weibos”—the Chi­nese ver­sion of tweets—are removed, often because they con­tain polit­i­cally sen­si­tive issues.

In research recently pre­sented at the inau­gural ACM Con­fer­ence on Online Social Net­works, Christo Wilson, an assis­tant pro­fessor in the Col­lege of Com­puter and Infor­ma­tion Sci­ence, and his grad­uate stu­dents Le Chen and Chi Zhang sought to answer two impor­tant ques­tions about cen­sor­ship on Sina Weibo: first, does the phe­nom­enon actu­ally stifle dis­cus­sions? And second, do the site’s users change their behavior in response?

The study rep­re­sented the first look at the impact of cen­sor­ship on users’ actual activ­i­ties on the social network—and what the researchers dis­cov­ered was rather sur­prising. “We observed no chilling effect,” Le said. “Instead, cen­sored topics see more active users tweeting more frequently.”

The team used com­pu­ta­tional methods to ana­lyze more than 800 mil­lion weibos. They found that more than three-​​dozen topics gen­er­ated lots of dis­cus­sion and were rel­e­vant to real-​​world events during the period between March 30 and May 13 of this year.

For example, the watch inci­dent gen­er­ated nearly 1,000 orig­inal tweets, 10,000 retweets, and 8,000 com­ments over six days. (As on Face­book, Weibo users can com­ment on posts.) Of all that activity, how­ever, 82 per­cent was removed through var­ious cen­sor­ship mech­a­nisms, which include thou­sands of crowd­sourced workers who man­u­ally examine each tweet. The watch inci­dent still went viral, even though the North­eastern research team found it to be among the top-​​five most-​​censored topics among the 37 total it examined.

Wilson and his team found that in the most cen­sored cases, Weibo users came up with “morphs,” or alter­nate forms of pre­ex­isting words or phrases to pre­vent their dis­cus­sions from being cen­sored. For instance, instead of using Fan Jiyue’s actual name, Weibo users wrote terms such as “Lushan sec­re­tary” and “brother watch print” to evade watchful eyes.

Another case involved a fake news story about Chi­nese Pres­i­dent Xi Jingpin taking a “common” taxicab. There, Wilson’s team iden­ti­fied a series of morphs, including words that have no rel­e­vance to the topic, but which sound or look sim­ilar when written with Chi­nese char­ac­ters. The team found an ini­tial spike in dis­cus­sions sur­rounding the taxi article when it was first revealed as a hoax, but it was quickly sti­fled by cen­sor­ship. At the same time, the morphs began to take over, gen­er­ating up to four times as much activity as the ini­tial tweets by the time the orig­inal word had fallen off the map entirely.

“Weibo is fun­da­men­tally dif­ferent from other social net­working sites,” said Wilson, pointing to the cen­sor­ship users face.