Sina Weibo, China's Twitter-like microblogging platform, has seen tremendous growth since launching in 2010, with the site's 300 million users combining to post 70,000 messages per minute. This growth has forced the Chinese government to ramp up its censorship tactics, though its precise methods have thus far remained a mystery. Researchers in the US, however, have now attempted to pull back the curtain on the country's operations, as part of a new study released this week.

Without any firm details on China's censorship practices, Rice University's Dan Wallach and his team of researchers instead attempted to reverse-engineer the government's techniques by tracking Weibo posts as they appeared in real-time. After following activity from 3,500 users over a 15-day period, they found that about 13 percent of all posts had been deleted. Some, of course, had been deleted by users themselves, but Wallach's real interest lay in those erased by third parties — identifiable by a unique "permission denied" message that would appear after deletion. It's these "system deletions," according to the authors, that provide the most accurate idea of how China's censorship machine actually operates.

Patterns point to a human hand at work

According to the study, five percent of all system deletions occur within eight minutes of posting, with the highest volume of deletions coming within the first five to ten minutes of posting. Overall, 30 percent of system deletions occur within half an hour, while 90 percent are completed within a day. This would suggest, then, that China's censors are monitoring Weibo activity in real-time, though it's likely that they deploy automated alerts for particularly sensitive keywords or phrases, such as "sex scandal," "government," or "politician."

Nevertheless, they estimate that the country's system would require about 1,400 censors at any given moment. Assuming each employee works an eight-hour shift, that would result in a total of 4,200 workers on the government's payroll on a given day. Further supporting their hypothesis on human-fueled censorship are unique patterns observed over 24-hour cycles. Wallach noticed, for instance, that deletions are less frequent at night, and that the system faces a backlog of objectionable posts each morning, presumably when most workers are coming into the office. They also noticed a slight downturn in activity around 7 PM, when China's national news airs every night.

It's a fascinating study, and one that Wallach and his team hope to expand upon going further. The next step will be to determine how China prioritizes the content it deletes.