A variational-autoencoder approach to solve the hidden profile task in hybrid human-machine teams
Online, social media bots have been accused to spread misinformation and support extreme or minority-held opinions. However, bots in hybrid human-machine teams can also be designed to improve team performance. In this paper, we study the effect of a single minority-supporting bot in hybrid teams in a carefully controlled experiment. People working in teams of 10 were asked to solve a hidden-profile prediction task, where task-relevant information was scattered unequally across team members. To do well in this task, pieces of information shared by the minority and the majority of players should be integrated. Simple majority-based decisions are not enough to perform well as information held by minority players is also valuable. We used a variational auto-encoder to train a bot to learn people's information distribution by observing how people's judgements correlated over time. After training, a bot was designed to increase team performance by selectively supporting opinions proportionally to their under-representation in the team. We show that the presence of a single bot (representing 10\% of team members) can significantly increase the polarization between minority and majority opinions by making minority opinions less prone to social influence. Although the effects on hybrid team performance were negligible, the bot presence significantly influenced team opinion dynamics. These findings show that unsupervised learning can be used to program bots that can improve team performance.