Artificial agents are becoming increasingly present in our daily environment. From vocal assistants to humanoid robots, we are observing a change in the role played by these new entities in our lives (Samani et al., 2013). However, it is still a matter of debate as to whether humans perceive embodied artificial agents, such as humanoid robots, as social and intentional agents or simple artifacts (Hortensius and Cross, 2018; Wykowska et al., 2016). Several researchers have investigated whether humans would deploy similar sociocognitive mechanisms when presented with a novel type of (artificial) interaction partner (i.e., humanoid robots) as they would activate in an interaction with another human (Saygin et al., 2012; Cross et al., 2019; Wykowska, 2020).

In this article, we report a study in which we investigated whether robot behavior—by being humanlike or mechanistic—can modulate the likelihood of people adopting the intentional stance (Dennett, 1971). The study also addressed the question of whether pupil dilation—a marker of cognitive effort—can predict the type of stance people would adopt toward the robots, and how all these factors are related to individual “mentalistically inclined” or “mechanistically inclined” biases.

According to Dennett (1971), the intentional stance is a strategy that humans spontaneously adopt to interpret and predict the behavior of other humans, referring to the underpinning mental states (i.e., desires, intentions, and beliefs). The intentional stance is an efficient and flexible strategy, as it allows individuals to promptly interpret and predict others’ behavior. However, when interacting with nonbiological systems, humans might adopt a different strategy, which Dennett describes as the design stance. According to the author, we deploy this strategy when explaining a system’s behavior based on the way it is designed to function. The intuition behind Dennett’s definition is that humans would adopt the stance that allows them to predict and interpret the behavior of a system in the most efficient way. Thus, the adoption of either stance is not predefined; on the contrary, if the adopted stance is revealed as inefficient, one can switch to the other stance.

Several authors have demonstrated that people tend to spontaneously adopt the intentional stance toward other human and nonhuman agents (Abu-Akel et al., 2020; Happé and Frith, 1995; Heider and Simmel, 1944; Zwickel, 2009; see also Perez-Osorio and Wykowska, 2019a and Schellen & Wykowska (2019) for a review). However, it is not yet entirely clear which of the two aforementioned stances humans would adopt when interacting with humanoid robots. On the one hand, humanoid robots present humanlike characteristics, such as physical appearance (Fink, 2012). Hence, it is possible that these characteristics elicit representations and heuristics similar to those that we rely on when interacting with humans (Airenti, 2018; Dacey, 2017; Waytz et al., 2010; Złotowski et al., 2015). This might trigger the neural representations related to the adoption of the intentional stance (Chaminade et al., 2012; Gallagher at al., 2002; Ozdem et al., 2017; Spunt et al., 2015). Indeed, the presence of humanlike characteristics is one of the key factors that, according to Epley et al., 2007, contribute to anthropomorphism toward artificial agents, facilitating the adoption of the intentional stance. On the other hand, humanoid robots are man-made artifacts, and therefore, they might evoke the adoption of the design stance, as they can be perceived simply as machines (Wiese et al., 2017).

Recent literature has addressed the issue of adopting the intentional stance toward robots. For example, Thellman et al., 2017 presented a series of images and explicitly asked their participants to rate the perceived intentionality of the depicted agent (either a human or a humanoid robotic agent). The authors reported that participants perceived similar levels of intentionality behind the behavior of the human and the robot agents. Marchesi et al. (2019) investigated the attribution of intentionality to humanoid robots, developing a novel tool, the InStance Test (IST). The IST consists of a series of pictorial “scenarios” that depict the humanoid robot iCub (Metta et al., 2010) involved in several activities. In Marchesi et al. (2019), participants were asked to choose between mentalistic and mechanistic descriptions of the scenarios. Interestingly, individuals differed with respect to the likelihood of choosing one or the other explanation. Such individual bias in adopting one or the other stance toward humanoid robots called for examining whether it is possible to identify its physiological correlates. In fact, Bossi et al. (2020) examined whether it is possible to relate individual participants’ EEG activity in the resting state with the individual likelihood of adopting the intentional or design stance in the IST. The authors found that resting-state beta activity differentiated people with respect to the likelihood of adopting either the intentional or the design stance toward the humanoid robot iCub. Recently, Marchesi et al. (2021) have identified a dissociation between participants’ response time and the stance adopted toward either a human or a humanoid robot. Moreover, the individual bias emerged as being linked to participants’ individual tendency to anthropomorphize nonhuman agents.

Since the literature presents evidence for various individual tendencies to adopt either the design or the intentional stance, in the present study, we aimed at using pupil dilation as a marker of individual bias and cognitive effort invested in the task of describing a robot's behavior, by adopting either stance. In addition, we were interested in finding out whether observing different types of robot behavior (humanlike or mechanistic) would have an impact on adopting the two different stances, taking into account individual biases.

In line with Marchesi et al. (2019), the score in the InStance Test was calculated ranging on a scale from 0 (extreme mechanistic value) to 100 (extreme mentalistic value). In order to obtain the average InStance Score (ISS) per participant, the scores across single scenarios were averaged. Before performing any preprocessing, the overall average score at the InStance Test after observing the mechanistic behavior was 43.80, with SD: 17.69, and the overall average score after observing the humanlike behavior was 43.44, with SD: 18.03 [t(65.97) = –0.08, p = 0.934]; thus, the type of robot behavior that participants observed did not modulate the ISS. The overall sample average score at the InStance Test was 43.62, SD: 17.26.

As in the study by Bossi et al. (2020), given that our focus was the individual bias at the IST, in the present section, we will report the results from the mechanistically and mentalistically biased participants, leading to an overall total sample of N = 21 participants. Results on the very same models involving unbiased participants as well are reported in the Supplementary Materials (overall N = 34 participants).

In the present study, we investigated whether adopting the intentional/design stance could be predicted by changes in pupil dilation and how both effects are modulated by participants’ individual bias in adopting the intentional stance and by a behavior of a robot observed prior to the test. To address these aims, we conducted an experiment in which participants first observed the embodied humanoid robot iCub, programmed to behave as if it was playing solitaire on a laptop positioned in front of it. From time to time, the robot was programmed to turn its head toward a second monitor on its left periphery, where a sequence of videos was being played. The behaviors exhibited by the robot were manipulated in a within-subjects design: in one condition, the robot exhibited a humanlike behavior, and in the second condition, the robot exhibited a machine-like behavior. After each session with the robot, participants’ pupil data were recorded while they completed the InStance Test. Participants were then divided into two groups, based on the bias showed by their IST score: a mentalistically biased group and a mechanistically biased group.

We found that both mechanistically and mentalistically biased participants leaned more toward mentalistic attributions in the IST after observing the robot’s humanlike behavior, as compared to the mechanistic behavior. This shows that participants had some sensitivity to the subtle differences in the robot behavior, thereby attributing more “humanness” to the humanlike behavior, independently of their initial bias (Ghiglino et al., 2020b).

We also explored the relationship between the individual bias and the changes in pupil dilation as a function of the behaviors displayed by the robot. We found that the two groups showed different patterns. On the one hand, for mechanistically biased people, pupil dilation was greater when they chose descriptions of the robot behavior in terms that were “congruent” with the previously observed robot behavior: a mentalistic attribution after the humanlike behavior and a mechanistic attribution after the machine-like behavior. We argue that this is due to the engagement of additional cognitive resources, caused by the cognitive effort in integrating the representation of the observed behavior into the judgment (Kool et al., 2010; Kool and Botvinick, 2014). In other words, these participants might have had enough sensitivity to detect the “human-likeness” or “machine-likeness” in the behavior of the robot. We argue that the integration of this piece of evidence into the judgment in the IST might have required additional cognitive resources.

On the other hand, mentalistically biased participants showed a tendency for greater pupil dilation when choosing the mentalistic description, independent of the observed robot behavior. Perhaps this group of participants showed engagement of additional cognitive resources when they were choosing descriptions that were in line with their initial bias (Christie and Schrater, 2015). Adherence to the “mentalistic” descriptions, independent of observed behavior, indicates, on the one hand, lower cognitive flexibility than the mechanistically oriented participants and, on the other hand, might be related to the general individual characteristic to structure and make the external world reasonable. This tendency to structure the external environment and engage in cognitive effortful tasks is defined as “need for cognition” (Cacioppo and Petty, 1982; Cohen et al., 1955; Epley et al., 2007). Mentalistically biased participants might have a lower need for cognition, and therefore pay less attention to all the subtle behavioral cues exhibited by the agent and stick to their original bias. Therefore, we may argue that this group is less prone to changing the stance adopted to interpret an agent’s behavior.

One last (and interesting) finding of our study was that RTs were faster on the extremes of the IST score distribution. This suggests that perhaps once participants made a clear decision toward mentalistic or mechanistic description, it was easier and more straightforward for them to indicate the extreme poles of the slider. On the other hand, when they were not convinced about which alternative to choose, they indicated this through keeping the cursor close to the middle and longer (more hesitant) responses.

Overall, it seems plausible that the general mechanistic bias leads to allocating a higher amount of attentional resources toward observation of the robot (Ghiglino et al., 2020a), resulting in paying more attention to the details of the observed behavior (in line also with Ghiglino et al., 2020b; see also Marchesi et al., 2020). This, in turn, might influence the subsequent evaluation of robot behavior descriptions. On the other hand, a mentalistic bias might lead participants to stick to their spontaneous first impression (Spatola et al., 2019) and a lower need for cognition (Cacioppo and Petty, 1982; Cohen et al., 1955; Epley et al., 2007). Commonly, individual differences and expectations shape the first impression about a humanoid robot (Ray et al., 2008, Bossi et al., 2020, Horstmann and Krä mer, 2019; Marchesi et al., 2021). Perez-Osorio et al. (2019b) showed that people with higher expectations about robots tend to explain the robot behavior with reference to mental states. This might indicate that our participants with a mentalistic bias were predominantly influenced by their expectations about the abilities of the robot and, therefore, paid less attention to the mechanistic behaviors of the robot. To conclude, we interpret the results in light of the influence of individual differences in the allocation of cognitive resources that might differ between people who are prone to adopting the intentional stance toward humanoid robots and people who, by default, adopt the design stance (Bossi et al., 2020; Marchesi et al., 2021).

In the present study, we opted for a within-subjects design to reduce the influence of interindividual differences related to prior knowledge/experience with the iCub robot. Nevertheless, we cannot rule out the fact that our approach was indeed too conservative, leading to a null effect of the robot behavior manipulation on the raw IST scores due to a carry-over effect. Future research should consider adapting similar paradigms to a between-subjects design, since this option will allow for controlling possible carry-over effects.

In conclusion, our present findings indicate that there might be individual differences with respect to people’s sensitivity to subtle hints regarding human-likeness of the robot and the likelihood of integrating the representation of the observed behavior into the judgment about the robot’s intentionality. Whether these individual differences are the result of personal traits, attitudes specific to robots, or a particular state at a given moment of measurement remains to be answered in future research. However, it is important to keep such biases in mind (and their interplay with engagement of cognitive resources) when evaluating the quality of human–robot interaction. The evidence for different biases in interpreting the behavior of a humanoid robot might translate into the design of socially attuned humanoid robots capable of understanding the needs of the users, targeting their biases to facilitate the integration of artificial agents into our social environment.