Media Responses to Government Repression: Machine Learning Evidence from Tanzania

• In Event: Authoritarian Propaganda & Censorship

Abstract

Across the world, aspiring and entrenched autocrats often seek to manipulate their countries' news environments. Given the global rise of autocrats willing and able to manipulate the informational environment, it is crucial to understand how different media sources and broader information ecosystems respond to government repression.

Previous research has shown that ownership affects the content of media outlets. Private ownership and a healthy stream of advertising revenues allow the media to become insulated from government pressure. On the other hand, the largest outlets are increasingly owned by conglomerates with ties to authoritarian regimes. Foreign-owned outlets have the potential to be resilient against attacks since they depend less on the government, yet they tend to be a primary target of legislation that restricts journalistic work through regulations of foreign ownership in the domestic media. Our understanding of whether and why some types of outlets and ownership structures are more resilient to attacks on media freedom is still in its infancy. This is particularly true in the case of alternative news outlets, such as online forums and blogs, even though these platforms can provide space for citizens to express themselves outside of the traditional media ecosystem.

We develop an original argument on how various types of media respond to the kind of restrictive legislation that has become so common during the current global trend toward autocracy. We argue, first, that restrictive legislation can have a distinct impact on what the news covers and on the sentiment with which it is covered. Second, we argue that the news media is conditioned by whether it is domestically or internationally owned. We hypothesize that international papers initially respond by focusing on the repressive behaviors of the government with negative sentiment, but their coverage of the country diminishes through time since international news sources tend to rely on national news sources to do much of the day-in-day out journalism. Third, we argue that domestic news sources that are initially close to the government will continue to cover similar issues albeit with even more pro-government sentiment. In contrast, formerly critical news sources will shift the substance of what they cover away from regime-relevant events and its coverage will have less critical sentiment. Fourth and finally, we argue that much of the journalistic substance and negative sentiment formerly expressed by critical national news sources will migrate to informal social network sites.

We test these claims using an enormous corpus of electronic media data and exploiting a significant legal change that restricted media freedom. We employ a state-of-the-art neural network model, Bidirectional Encoder Representations from Transformers (BERT), to classify whether articles cover events bearing on the regime and civic spaces as well as the sentiment (i.e. pro- or anti-government) of the coverage. Our data consists of daily news events published by 14 different media outlets, and posts from a popular internet forum during the period 2014 to 2020. We exploit the major restrictive legal changes in 2016 and 2018 to uncover their heterogeneous effects on news reporting in the country. We focus on two aspects of the news coverage: content --i.e., the kinds of events the news outlets cover-- and sentiment --i.e., the degree to which reporting becomes biased in favor of the government. Our data allow us to assess changes in news coverage across a wide variety of media news outlets, from international news organizations based in the developed world and other African countries to national newspapers and reputable online news forums. Our results will shed light on the resilience of news organizations in the face of advancing democratic backsliding across the world.

Authors

• Fatih Serkant Adiguzel, Duke University

• Diego José Romero, Duke University

• Erik M. Wibbels, Duke University