A Political Model of Trust

Abstract : We analyze a model of political competition, in which the uninformed median voter chooses whether to follow or ignore the advice of the elite that forms endogenously to aggregate information. In equilibrium, information transmission is possible only if voters trust the elite's endorsement of potentially biased candidates. When inequality is high, the elite's informational advantage is minimized by the voters' distrust. When inequality reaches a certain threshold, the trust, and thus the information transmission, breaks down completely. Finally, the elite size and thus the extent of information aggregation depends on the amount of trust they can maintain.

---

A Model of Censorship, Propaganda, and Repression

Abstract : We extend the canonical two-state, two-action model of Bayesian persuasion to explore the interaction among censorship, propaganda, and repression in autocracies. Censorship renders propaganda (persuasion) more effective but blocks information useful to the government in deciding whether to repress. When the government has the capacity to implement any censorship level with precision, propaganda is uninformative in equilibrium; repression is a last resort when censorship fails. When the government instead implements censorship with error, propaganda is informative when the cost of repression is high. Our analysis highlights that information manipulation and repression may occur in tandem and that the option to repress affects the nature of censorship and propaganda.

---

Inauthentic Newsfeeds and Agenda Setting in a Coordinated Inauthentic Information Operation

Abstract : The 2015-2017 Russian Internet Research Agency’s coordinated information operation is one of the earliest and most studied of the social-media age. A set of 38 city-specific inauthentic “Newsfeeds” made up a large, under-analyzed part of its English-language output. We label 1000 tweets from the IRA Newsfeeds and a matched set of real news sources from those same cities with up to five labels indicating the tweet represents a world in unrest and, if so, of what sort. We train a natural-language classifier to extend these labels to 268k IRA tweets and 1.13m control tweets. Compared to the controls, tweets from the IRA were 34 percent more likely to represent unrest, especially crime and identity danger, and this difference jumped to about twice as likely in the months immediately before the election. Agenda-setting by media is well known and well- studied, but this weaponization by a coordinated information operation is novel.

---