News Archives

UNM, Rice Researchers Document the Velocity of Censorship

March 13, 2013

Jed CrandallCensors at Sina Weibo, a Chinese website similar to Twitter, work with amazing speed and efficiency. That’s the conclusion of research conducted by UNM Assistant Professor of Computer Science Jed Crandall (in photo) and Rice University Professor of Computer Science and Electrical Engineering Dan Wallach in conjunction with an independent researcher and an undergraduate researcher from Bowdoin College. The study is titled, “The Velocity of Censorship: High-Fidelity Detection of Microblog Post Deletions,” and is undergoing peer-review.

“The point of our measurement study of Weibo is to take a closer look at global online censorship practices,” Crandall said. “There has been considerable debate in the U.S. recently about extending copyright law enforcement to include various kinds of filtering online. China already has laws in place for companies within China to filter online content.”

Weibo is one of the biggest social network companies in China, and it faces the dual challenge of keeping its users engaged (and thus, watching advertisements and making money for Weibo) while keeping the content it hosts compliant with local laws. If Weibo had insufficient controls, the government may take action against the company. If their controls were too rigid, users might abandon them for one of their competitors. Weibo’s success implies that it has found a happy medium, and the research team says that is what makes Weibo an interesting social media platform to study.

In February 2012 Weibo had more than 300 million users and about 100 million messages that were sent daily. Weibo, like Twitter limits message length to 140 characters. It also allows embedded photos and videos and comment threads to be attached to posts.

In spite of the tremendous volume, the company can detect a censorship event within one minute of posting. In their paper the researchers describe how they were able to track the censors at work and hypothesize several different filtering methods that appear to comprise Weibo’s defense-in-depth system of censorship.

“Weibo gives us a window into the future for what Internet censorship of social media around the world may look like,” Wallach said. “Former Supreme Court Justice Louis Brandeis championed transparency a century ago when he wrote, ‘sunlight is said to be the best disinfectants.’ We hope that our research shines a light on how laws created by governments and implemented by the private sector can affect free speech everywhere, including here in the U.S.”

One of the most interesting elements of the research is the finding about the speed at which the censorship process actually works. According to Crandall, “There have been some studies on Weibo showing that posts are deleted after a day or two, but we see posts being deleted after five or ten minutes. Basically we showed that if you want to have a complete picture of internet censorship you have to have something that can measure very quickly on the order of minutes, and you have to be able to measure a wide variety of things.”

About the Research

The research group first had to determine how to approach the problem, so they looked at the postings of individuals who had previously been censored. Over weeks, they were able to slowly expand their group, adding any user with more than five deleted posts and eventually finding more than 3,500 users to track for their sample.

Once they settled on a group to monitor, they decided to check their posts every minute to determine how long it took censors to remove a post. Their web crawler searched for posts that appear, and then are subsequently deleted. Their data showed that five percent of the deletions happened in the first eight minutes and within 30 minutes nearly 30 percent of the deletions were completed. More than 90 percent of the deletions occurred within one day after a post appeared.

This information lead the team to think about what resources it would take to monitor that river of information. They calculated it would take 4,200 workers reading 50 posts a minute in eight hour shifts to censor using only human review of the posts. That led them to the conclusion that much of the filtering must be automated, through initial flagging and retrospective searches. This conclusion allowed them to set up six hypotheses.

There is a surveillance keyword list that triggers for posts to be look at by a moderator for possible deletion.

Weibo targets specific users, such as those who frequently post sensitive comments.

When a sensitive post is found, a moderator will find all of its related reposts and delete them all at once.

Weibo removes posts retroactively via keyword search, causing spikes in the deletion rate of a particular keyword within a short amount of time.

The censors work relatively independently, in a distributed fashion. Some of them may work in their spare time.

Deletion speed is related to the topic. That is, particular topics are targeted for deletion based on how sensitive they are.

The research team notes there may be many mechanisms beyond those they have hypothesized, which future work may reveal. The team did not consider interactions between social media and traditional media but suggest that would be an interesting topic for future research.

Media contacts:

UNM Karen Wentworth (505) 277‑5627; email: 
Rice University, Jade Boyd (713) 348‑6778; email:
David Ruth (713) 348‑6327; email: