Chapter 10: Social Cognition And Human-Robot Interaction

The diffusion of new technologies and smart devices is giving researchers new tools for scientific inquiry in the areas of experimental psychology and cognitive sciences (Wykowska, 2021). Nowadays, even “pocket technologies” seem to display a certain degree of “artificial thinking,” which might remind us of human mental processes (Ghiglino & Wykowska, 2020). This similarity could affect the human tendency to attribute anthropomorphic traits to technologies, evoking social cognition mechanisms that are typical of human-human interaction. However, it is still an open question whether it is possible (or even desirable) to ascribe a “mind” to artificial agents. Several authors claimed that attributing mental states might help individuals in predicting and explaining the behavior of artificial agents, which, in turn, should positively affect collaborative scenarios (de Graaf & Malle, 2019; Imamura, et al., 2015; Levin, et al., 2013; Thellman, et al., 2017). Addressing this issue seems to be pivotal for fundamental sciences as well as for the design and development of artificial agents that are supposed to assist humans in the everyday life of the future. Indeed, this question has been approached already in the 1970s, when the philosopher Daniel Dennett proposed a model to explain how humans process the behavior displayed by agents of a different nature (Dennett, 1971). Dennett (1971) hypothesized that humans adopt spontaneously different strategies when interacting with inert objects, complex artifacts, and biological systems. For example, when humans need to explain and predict the behavior of a simple object, such as a ball rolling on an inclined plane, they tend to rely on physical laws guiding the behavior itself (i.e., gravity pull, attrition, etc.). Thus, under such circumstances, humans tend to adopt what Dennett calls the “design stance” (p. 88). However, if the object displaying the behavior is more unfamiliar and complex than a rolling ball, such as a car approaching a crossing, this stance might not be adequate to provide the perceiver with reliable explanations and predictions. In these cases, humans need to rely on more complex representations and refer to the design and functionality of the artifact displaying the behavior, adopting what the Dennett calls the design stance. Indeed, the fact that the car is moving is not explicable solely in terms of inertial motion, but it becomes clearer as soon as we start thinking about the design of the motor or the brake. However, there are entities whose behavior can be summarized neither as a mere effect of physical laws nor as a consequence of their design, as in the case of biological agents. Indeed, if we talk with a person that is cooking, and they repetitively move their head in the direction of the oven, we need to include in our predictive model a representation of the agent’s internal states (in this case, the inference that the person wants or needs to check if the food in the oven is at the right temperature) adopting what Dennett (1971) calls the “intentional stance” (p. 90). This means that we need to treat the action as motivated by the agent’s mental states (i.e., beliefs, desires, feelings, and intentions). In other words, the intentional stance consists of treating the behavior as a consequence of mental operations, which are directed towards a goal. Importantly, the adoption of the intentional stance does not imply the ascription of a mind to the agent that is displaying the behavior (Dennett, 1981). On the contrary, its function is limited to providing the most accurate strategy available to approach an agent whose behavior is ambiguous, even when the agent is artificial (Marchesi, et al., 2019). This implies that under certain circumstances, humans might adopt the intentional stance even when interacting with complex artificial agents. It has been found, for example, that humans often describe complex artificial agents (i.e., robots) using a mentalistic vocabulary, referring to their intentions or will, while also denying that they have a mind (Banks, 2020). Indeed, the ambiguity of an agent’s behavior, especially if the agent is human-like, might recall cognitive processes that are normally recruited during the interaction between biological agents. We can even argue that the tendency to attribute human traits to artificial agents might extend to social cognition mechanisms in general, which can be evoked spontaneously when interacting with technologies that resemble human beings. Indeed, social cognition strategies are flexible, adaptable, and depend upon the knowledge, experience, and representation that an individual develops around a certain agent and/or context (Wykowska, 2020). Fifty years after Dennett’s first model, the scientific debate on social cognition mechanisms that are evoked during the interaction between humans and artificial agents is still heated. In particular, robotic technologies are of great interest, because of their embodiment and their physical presence, which allow them to actively interact with their environment. Furthermore, the possibility to manipulate objects, jointly with humans, makes such platforms suitable to investigate social cognition mechanisms in naturalistic scenarios (Moreno & Etxeberria, 2005). Yet, artificial agents can still be programmed to behave exactly as designed, offering excellent experimental control for research paradigms. In recent years, human-robot interaction (HRI) has become an interdisciplinary trend, attracting social scientists as well as engineers, interested in understanding whether the behavior of robots can be perceived by human partners as a reflection of mental operations (Frith & Frith, 2008). Such diversity requires inter- and transdisciplinary discussions to define research methods that are suitable to investigate social cognition applied to HRI (Thellman, et al., 2021). In the following sections of the chapter, we will present some key concepts of HRI research, trying to define a common ground for practitioners. We will first explore the concept and definition(s) of embodiment, as it may be considered one of the crucial aspects of contemporary HRI, offering researchers the possibility to experiment with the second-person neuroscience approach¹ (see Schilbach, et al., 2013). Then, we will briefly present the most common methodologies applied to HRI, highlighting the complementarity of different approaches. Eventually, we will present some of the practical implications of multidisciplinary HRI.