Proceedings of the Annual Meeting of the Cognitive Science Society

In deception games, deceivers must find ways to draw in unknowing bystanders, and bystanders must develop strategiesfor detecting falsehoods. What are the strategies that people use in these roles, and can computer systems also detect thesebehaviors? We address this question through text-based games of Mafia, wherein players are assigned to deceptive roles(mafia) or roles incentivizing detecting deception (bystanders). We find that participants adopt sophisticated role-basedstrategies, wherein the mafia, who are outnumbered but know the identities of all players, act carefully to secure the votesof the bystanders by speaking more even as verbose speakers tended to be eliminated. These role-based behaviors weredistinct enough that a computational classifier could distinguish between mafia and bystanders with 70.3% accuracy andoutperform human players. Understanding the systematic features defining honest and deceptive players advances ourability to automatically detect online deceit and grasp group dynamics in real-world collaboration.