PUB - Publikationen an der Universität Bielefeld

Humans are capable of what is commonly referred to as mind-reading or Theory of Mind, meaning that they are able to infer the mental state of other agents based on observable cues such as their behavior. This is a vital skill for human-human interactions, both for cooperation as well as competition, since it allows us to better predict other agents’ future behavior. While one may think that humans always include as much information about other agents as possible in order to maximize their prediction’s accuracy, we find that humans are quite flexible in how much effort they exhibit when mind-reading. We found sub-optimal inference of another agent’s mental state in earlier work (P¨oppel & Kopp, 2019), where participants projected their own knowledge onto the observed agent despite having seen contradicting evidence, i.e. a strong egocentric bias. This effect was especially noticeable, when participants were not primed to actively consider certain possible mental states of the agent.We were able to show that ”satisficing mentalizing” can improve prediction accuracy while reducing computational costs in artificial systems by only switching to more costly mentalizing approaches, when simpler heuristics failed to predict observed behavior (Pöppel & Kopp, 2018). These results let us to believe that humans similarly employ ”satisficing mentalizing”: While capable of highly complex mental inferences, they tend to fall back to simpler explanations or heuristics in many situations.We understand these findings in that they conform to the notion of bounded rationality (Simon, 1955; Lieder & Griffiths, 2019). Making complex inferences incurs costs in the form of mental effort. However, it is currently unclear what factors actually influence the amount of mentalizing effort humans are investing in different situations. In order to uncover some of these factors involved in making predictions, we performed a series of experiments manipulating different variables, such as the environment’s complexity or the costs of wrong predictions for an observed, virtual agent. Preliminary results across different experiments replicate various degrees of heuristics, often but not exclusively in the form of egocentric biases, participants appeared to use when predicting the observed agents’ behavior.We also found evidence that higher costs for wrong predictions as well as priming participants to consider mental states reduce the egocentric bias. Furthermore, regardless of the performed mentalizing, participants used attributed mental states, primarily the agent’s desire, in order to explain their choices afterwards. When people use mental states attributed to artificial systems when reasoning about their own decisions, while not always performing optimal mentalizing, it raises the question how artificial systems should ideally react, especially to potential changes in the employed mentalizing. Can we design robots and/or their behavior in such a way as to elicit more complex mentalizing? We are convinced that in order to significantly improve HRI, and to leap to the next level of human-robot intersubjectivity, will require us to unravel these fundamental socio-cognitive processes, including the involved factors for mode switching, and to consider them in building robots that need to understand and be understood by humans on a more deeply social level.