Strategically Misleading the User: Building a Deceptive Virtual Suspect

Humans lie every day, from the least harmful lies to the most impactful ones. Therefore, in an attempt to design virtual agents endowed with advanced decision-making abilities, researchers not only focused their effort in designing cooperative and truthful agents but also deceptive and lying ones. In this paper we propose a model capable of engaging an agent in an uncooperative misleading dialogue with a user. This model gives to an agent the ability to reason about its knowledge and then autonomously adjust the story it tells depending on what its interlocutors might know and on how sensitive it considers the conversation topic to be. Such a model allows a story's author to focus on the main narrative, letting the model handle the generation of alternatives. We implemented the model in an agent called the Deceptive Virtual Suspect and conducted some preliminary experiments using an Interrogation Game.