Known Unknowns: An Analysis of Twitter Censorship in Turkey.

ABSTRACTTwitter, widely used around the world, has a standard interface for government agencies to request that individual tweets or even whole accounts be censored. Twitter, in turn, discloses country-by-country statistics about this censorship in its transparency reports as well as reporting specific incidents of censorship to the Chilling Effects web site. Twitter identifies Turkey as the country issuing the largest number of censorship requests, so we focused our attention there. Collecting over 20 million Turkish tweets from late 2014 to early 2015, we discovered over a quarter million censored tweets - two orders of magnitude larger than what Twitter itself reports. We applied standard machine learning / clustering techniques, and found the vast bulk of censored tweets contained political content, often critical of the Turkish government. Our work establishes that Twitter radically under-reports censored tweets in Turkey, raising the possibility that similar trends hold for censored tweets from other countries as well. We also discuss the relative ease of working around Twitter's censorship mechanisms, although we can not easily measure how many users take such steps.