A Study of AI Agent Commitment in One Night Ultimate Werewolf with Human Players

Social deduction games are a genre of board games in which a group of players is secretly assigned roles and each player tries to determine the other players’ roles. However, some roles have an incentive to not be found, and the games typically allow players to lie freely. Playing such games is a challenging task for AI agents, because they need to not only determine the probability that each statement made by the other players is truthful, but also come up with convincing lies themselves. In this paper, we present AI agents designed to play one particular such game, One Night Ultimate Werewolf, with human players. We discuss the different deliberation strategies our agents use to determine what they should say, and when they should change their plan. To determine how these different deliberation strategies are perceived by human players, we performed an experiment in which participants played a Unity implementation of the game with each of the three deliberation strategies. We present the results of this experiment, which show that commitment to plans has a measurable effect on player perception and provide a trade-off between consistency and potential for high performance of the agent.