Dynamics of Polarization on Online Platforms and AI Remedies

Online platforms connect individuals from diverse backgrounds across the globe, fostering a sense of global community. However, this vast reach also carries a potential risk: the formation of echo chambers that inadvertently reinforce existing stereotypes and contribute to polarization. This project uses the revision history of the collaboratively edited wikiHow1 platform to investigate the extent to which such a risk can be evidenced by measurable dynamics and how artificial intelligence methods can be used to mitigate polarization. Specifically, we study instructional texts written in a version for women and a version for men and examine how audience-specific changes over time contribute to making articles more polarizing and in how far such changes actually suit the needs and preferences of the corresponding target groups. Finally, we test whether large language models (LLM) can remedy undesired polarization effects by merging articles written for specific audiences into more balanced articles for a general audience. We bring together methods from computational linguistics and experimental psycholinguistics to carry out the project and to measure its success.