Homophily in Social Networks and the Propagation of False Information
Jonas Stein, Arnout van de Rijt, Marc KeuschniggPast studies using observational data and simulation have suggested that echo chambers (homophilous segments of online networks) boost the spread of false information. However, it is hard to control for factors that may impact the spread of false information when using observational data available on social platforms such as Facebook or Twitter. Therefore, this project takes a new approach by isolating the effect of network homophily through an experimental setting, enabling a controlled test of whether homogenous networks facilitate the dissemination of false information.
From a theoretical standpoint, one would not necessarily assume that only false information would spread widely in echo chambers. In fact, stronger social influence and confirmation bias should help both true and false messages to perform better in an ideologically aligned homogenous network. In order to explain why past empirical evidence stresses the significance of echo chambers specifically regarding false information, we propose a new theoretical argument: Whereas the diffusion of true messages is facilitated by higher general credibility in any network, false messages will only surpass a percolation threshold and diffuse widely if social influence and confirmation bias effects from a homogenous network are present.
In addition, it is addressed whether such hypothesized dynamics can be offset by incentivizing people to share true information only. Two competing expectations exist in this case: Either, incentives may be able to eliminate increased percolation of false messages in homogenous networks, suggesting that greater echo chamber effects can be diminished if people are encouraged sufficiently. Alternatively, incentives may have little or no influence, suggesting that it is either network member’s lack of ability to identify messages as false, even when incentivized, or network members are willing to pay for voicing their standpoint, despite knowing that the standpoint is factually wrong.
The study uses experimental online social networks in which a total of approximately 4,000 liberal or conservative US American study participants can receive and share news messages. News messages are either true or false, and targeted to appeal to either liberals or conservatives. We vary the network structure (heterogeneous versus homogeneous) and participant payoff (flat-fee versus incentive to share true messages only). This results in a 2x2 experimental setup, enabling us to test and compare the performance of given messages in each network condition.