Intergroup Conflict and Human-Algorithm Interactions
Repository URI
Repository DOI
Change log
Authors
Abstract
How do human-algorithm interactions shape intergroup conflict? This thesis combines psychological theory with methods from computer science to examine social identity biases in widely used algorithmic systems. I focus on investigating how ingroup solidarity and outgroup hostility are propagated by social media and large language models in highly polarised contexts. Although prior work highlights the viral potential of divisive language on social media, including negative emotions and outgroup hostility, it largely overlooks the role of ingroup solidarity in intergroup conflict. To address this research gap, I study social media content during naturally occurring intergroup conflicts in Ukraine and Israel, utilising advances in language modelling to measure how both ingroup solidarity and outgroup hostility predict engagement over time. Analysing over one million Ukrainian posts around the 2022 Russian invasion, I find that ingroup solidarity became a key predictor of engagement following the invasion. In a preregistered study in Israel after the October 7 Hamas attack, I further find that external threats can bridge internal political divides, shifting social media engagement patterns towards national ingroup solidarity and depolarising repost networks. I then conduct a preregistered longitudinal field experiment on Ukrainian Telegram to causally test the effects of group-aligned emotional news framing on social media engagement. Extending beyond social media, I develop a new methodology to examine how generative language models can inherit ingroup solidarity and outgroup hostility from human training data. Overall, the empirical results presented in this thesis highlight the importance of ingroup solidarity in intergroup conflict and the ways in which algorithmic platforms and systems learn and propagate human biases, posing new challenges for technological alignment.
Description
Date
Advisors
Roozenbeek, Jon
