In April 2014, an attack at the main railway station in Urumqi, a city in the northwest Chinese province of Xinjiang, killed three people and injured dozens more. The incident—an explosion followed by a knife attack—came at the end of President Xi Jinping’s first visit to the restive region since he took office, during which he had promised to ramp up the government’s response to terrorism.

Immediately after, the Chinese government’s online-censorship apparatus sprung into action. Searches for “Urumqi blast” were blocked on Baidu, the country’s largest search engine, and on Sina Weibo, a Twitter-like social network that’s very popular in China. Meanwhile, paid government trolls flooded Sina Weibo and various other Chinese social networks with more than 3,000 posts, in a coordinated burst of activity.

But, curiously, the posts had nothing to do with what had just happened in Urumqi. They didn’t start contentious debates, or push back against political arguments. Instead, they waxed poetic about China’s good governance, economic opportunities for Chinese people, and the “mass line.”

They were distractions. Rather than engage with chatter around the attack in Xinjiang, or try to tamp down political expression, the posts seemed to be designed to derail the conversation. Other big events—riots in another part of Xinjiang province in July 2013, for example, or a pair of important political meetings in February 2014—were met with similar spikes in government-sponsored social-media activity.

The coordinated volleys of posts are described in a research paper that will be published this year in the American Political Science Review, in which three scholars—Harvard’s Gary King, Stanford’s Jennifer Pan, and Margaret Roberts at UC San Diego—analyzed tens of thousands of posts written by China’s official social-media trolls.

The fact that the government coordinates friendly social-media posts isn’t new: The people behind them are known colloquially as the “50 Cent Party,” for the rumored sum they’re paid for each post. (They’re not actually members of a political party.) But by studying a large archive of emails leaked from one of the internet propaganda offices in Ganzhou—a city located in China’s southeastern Jiangxi province—the researchers were able to piece together details about how the operation works, and make some inferences about the campaign’s motivations.

Before setting out to research the 50 Cent Party, the scholars largely subscribed to the conventional wisdom among academics and journalists: that ordinary citizens were enlisted to debate with rabble rousers, and take a hard pro-government stance. Instead, they found very few instances of 50 Cent posts engaging in a back-and-forth, and a large volume of innocuous “cheerleading” posts that simply express goodwill about the government and its policies. Criticism on social media is largely tolerated, they found, but as soon as the risk of mobilization and collective action begins to loom, the government jumps in to disrupt the conversation.

Throughout the project, the researchers enjoyed unusual access to the inner workings of the Chinese government. The emails leaked from the Zhanggong propaganda office, for example, included the text of more than 43,000 50 Cent posts, sent from commenters proving that they’d completed their assignments, and messages from the propaganda office to higher-level offices.

The commenters themselves, it turned out, were nearly all identifiable as government workers. (There’s no evidence, the researchers said, that the government used bots to amplify its message.) They worked in various offices and bureaus, and didn’t appear to be paid at all for the posts. It may be that they were simply expected to post coordinated messages as part of their government jobs.

In emails from the propaganda office, commenters were instructed to “promote unity and stability through positive publicity,” and to “actively guide public opinion during emergency events”—where “emergency events” refer to events that might stoke collective action.

Based on what they found in the leaked archive, the researchers extrapolated the scale of the propaganda operation to the rest of China, estimating that the loosely-defined 50 Cent Party posts a total of 448 million messages on social media every year. “If these estimates are correct, a large proportion of government web site comments, and about one of every 178 social media posts on commercial sites, are fabricated by the government,” the researchers wrote.

To check their assumptions, the research team did something unusual: They reached out directly to social-media users they suspected of being 50 Cent party members. To make sure they’d developed a good model of telling 50 Cent posts from organic posts, they simply asked, kindly and in in Chinese, “I saw your comment, it’s really inspiring, I want to ask, do you have any public opinion guidance management, or online commenting experience?”

When they asked people they knew to be in the 50 Cent Party (because their information had turned up in the email leak) 57 percent admitted to being a part of a government operation. When they asked suspected 50 Cent members, 59 percent confirmed their status. Since the difference between the proportions was not statistically significant, the researchers concluded that they were accurately able to guess which posts belonged to 50 Cent commenters.

Even more unexpectedly, the researchers got semi-official confirmation that their findings were largely correct from the Chinese government itself. After an early draft of the research attracted media attention last May, the Global Times, a newspaper with strong ties to the government, published an editorial about it. The piece, which was written only in Chinese, defended the practices that the study uncovered: “Chinese society is generally in agreement regarding the necessity of ‘public opinion guidance,’” it read. (A quick study of social-media responses didn’t support the editorial’s conclusion: While only 15 percent of comments on the newspaper’s website were critical of “public opinion guidance,” 63 percent of comments on Weibo were disapproving.)

The piece also didn’t dispute any of the material points in the research, nor did it seek to discredit the contents of the leaked email archive. “For all practical purposes, the editorial constitutes the answer to a simple sample survey question,” the researchers wrote. “That is, instead of asking [50 Cent Party] members about their status as we [did earlier], we (inadvertently) asked the Chinese government whether they agreed with our results, and they effectively concurred.”

The distraction tactics that China’s troll army favors could just as easily be deployed elsewhere. Debating detractors directly, or censoring them outright can backfire—political scientists have found that more repression is sometimes correlated with increased mobilization—but derailing the conversation or diluting the intensity of collective criticism can be a sneaky way to defuse it.

“The activity and especially the scale we discovered in China was surprising and even shocking to us (and others), but the general strategy is not unique. You can find aspects of it on smaller scales from elected officials, corporation heads, citizens, and numerous others,” wrote Gary King, one of the researchers behind the study, in an email. “Usually it is not the official policy of a government.”