What the Censors Actually Delete

In 2013, Gary King, Jennifer Pan and Margaret Roberts published a study in the American Political Science Review that did what almost nobody had managed: it measured what Chinese censorship actually removes.

The method is the reason the finding holds. They collected millions of social media posts from nearly 1,400 Chinese services at the moment of publication, before the censors had reached them, and then returned to see which had disappeared. Nobody was asked what the rules were, because nobody would have said.

The result contradicted the sensible expectation.

Posts criticising the state, its leaders and its policies were not more likely to be deleted than posts that did not. Criticism, including sharp criticism of named officials, survived at ordinary rates. What predicted deletion was collective action potential: whether a post represented, reinforced or could spur people to assemble, protest or organise.

And the finding has a detail that fixes the interpretation beyond argument. Posts supportive of the government were also removed when they had that potential. A rally in favour of a policy is still a rally.

So the system is not thin-skinned. It is strategically indifferent to being disliked and extremely alert to being gathered against.

Once you see it, the logic is almost boringly rational. A government without competitive elections has a genuine information problem: it needs to know which local officials are failing, and the people with that knowledge have no ballot with which to report it. Complaint on social media is an inspection system that costs nothing to operate. Deleting it would be destroying the only instrument that tells the centre what is going wrong in a county it has never visited.

Coordination is a different matter entirely, because coordination is the thing that removes governments.

The same researchers confirmed the finding in 2014 by a method that is rare in this field: they set up their own social media site inside China, submitted content, and observed the rules from the position of an operator rather than a user. Very few claims about closed systems have that quality of confirmation behind them.

They also studied the paid posting operation usually called the fifty cent party, and found something that again cut against expectation. The volume is enormous, in the order of hundreds of millions of posts a year, and the content is overwhelmingly not argument. It is cheerful, patriotic, off-topic distraction. The operation is not trying to win exchanges; it is trying to change the subject, which is cheaper and works better.

The reason this chapter sits at the centre of the book is methodological as much as political. Everything here was established by measuring what disappeared rather than by asking an institution to describe itself. That technique generalises to every closed system in this subject, and it is the only reliable answer to the difficulty raised in the very first book: that the evidence tends to come from whoever is willing to hand it over.