Humans rely on non-verbal behavior to interact with one another. For example, when we observing someone looking at an apple, we infer that that person is hungry. This process, in which people make inferences about what others are thinking, is called mentalizing (Baron-Cohen, 1995), and is essential for successful human social interactions (Frith and Frith, 2006). Mentalizing is a highly automatic process that we engage in (Frith and Frith, 2006). Assuming that others' behavior is based on their thoughts, emotions, and internal states is called adopting the intentional stance (Dennett and Haugeland, 1987; Dennett, 1989). While it is assumed that when humans interact with one another they adopt the intentional stance (i.e., they explain and predict others' behaviors with reference to their underlying thoughts, feelings, and intentions), an interesting question remains if (and when) people adopt the intentional stance toward artificial agents that resemble humans, such as humanoid robots. Although prior studies (e.g., Marchesi et al., 2019) have uncovered instances where people do adopt the intentional stance toward humanoid robots, here we ask if certain procedures used in experimental design can actually create obstacles for measuring the intentional stance in human–robot interaction (HRI).

Several studies do in fact show that people are capable of adopting the intentional stance toward artificial agents in some cases. For example, when robots, avatars, or computers perform unexpected actions (Morewedge, 2009; Waytz et al., 2010), resemble humans in their physical appearance (Kiesler et al., 2008; Admoni et al., 2011), display emotions (Fussell et al., 2008), react to people's actions (Terada et al., 2007), exhibit humanlike non-verbal behaviors such as shrugging (Carter et al., 2014), cheat (Short et al., 2010), or engage in eye contact with others (Ito et al., 2004; Yonezawa et al., 2007), humans are more likely to adopt the intentional stance toward them. When humans adopt the intentional stance, it can lead to positive outcomes in HRI¹ as robots that are perceived as intentional are seen as more trustworthy (Kiesler et al., 2008; Carter et al., 2014), are able to engage humans better (Yamazaki et al., 2010), and have positive effects on people's moods (Carter et al., 2014). Adopting the intentional stance toward robots can also have positive effects on human performance in HRI as they can facilitate learning (Brown and Howard, 2013), improve physical interactions where hand–eye coordination is needed (Carter et al., 2014), induce social facilitation effects (Bartneck, 2003; Woods et al., 2005; Looije et al., 2010), and improve interactions in team settings (Breazeal et al., 2005). By the same token, not adopting the intentional stance toward social robots can pose a problem for HRI, as humans might be engaging fewer socio-cognitive mechanisms in social interactions with robots, and thereby less social attunement (i.e., activation of socio-cognitive mechanisms) occurs between the two agents (Wiese et al., 2012, 2018; Wykowska et al., 2014; Özdem et al., 2016; Caruana et al., 2017; Ciardo et al., 2020; for a review, see Perez-Osorio et al., 2015; Wiese et al., 2017; Schellen and Wykowska, 2019).

One successful paradigm to measure social attunement in HRI is the gaze-cueing paradigm (Wiese et al., 2012; Wykowska et al., 2014; Perez-Osorio et al., 2015; Kompatsiari et al., 2018) as prior work has shown that attentional orienting to gaze cues is the foundation of many (often higher-level) social-cognitive processes, such as mentalizing (Baron-Cohen, 1995; Nummenmaa and Calder, 2009; Teufel et al., 2010; Pfeiffer et al., 2011). The gaze-cueing paradigm generally starts with a gazing stimulus (i.e., a face or face-like drawing) that looks directly at an observer (i.e., participant) and then makes a gaze shift to a certain direction (e.g., right or left), which shifts the observer's attention to the gazed-at location (Friesen and Kingstone, 1998; Driver et al., 1999; Langton and Bruce, 1999; Quadflieg et al., 2004), if the observer engages in joint attention with the gazer. After the gaze- shift, a target appears either in the same (i.e., valid trial) or a different location (i.e., invalid trail) relative to the gaze cue. Studies have consistently shown that people are faster at responding to valid trials compared to invalid trials, with the magnitude of the difference between valid reaction times and invalid reaction times indicative of how strongly or how frequently attention was oriented to the gazed-at location (i.e., the gaze-cueing effect: GCE). Studies investigating the GCE have also shown that the effects are observed even when gaze was counter-indicative of the target location, which suggests the reflexivity of attention orienting to gaze cues (Ristic and Kingstone, 2005; Kingstone et al., 2019).

More recent studies have shown that attention orienting to gaze cues can be top-down modulated (i.e., volitional), depending on the social relevance of the gaze. For example, attention orienting to gaze cues is stronger when the gazer is of higher social rank (Shepherd et al., 2006; Jones et al., 2010; Ohlsen et al., 2013; Cui et al., 2014; Dalmaso et al., 2014), is physically similar to the observer (Hungr and Hunt, 2012), is perceived as trustworthy (Süßenbach and Schönbrodt, 2014), has the possibility to see (Teufel et al., 2010), or when gaze cues are meaningful to the context of the gaze (Perez-Osorio et al., 2015). Paramount to our study, prior work has also shown that GCEs are enhanced when gaze shifts are thought to originate from an intentional agent (Wiese et al., 2012; Wykowska et al., 2014; Caruana et al., 2017), compared to just a pre-programmed machine, which suggests that adopting the intentional stance toward a robot can influence how strongly it can induce attentional orienting to its directional cues.

Although previous literature showed that, indeed, people are able to socially attune with a robot, and perhaps more strongly so if they adopt the intentional stance toward it, more recent studies have shown that using artificial agents in gaze-cueing studies can have negative consequences (Abubshait et al., 2020a). Therefore, it is unclear if classical psychological paradigms such as the gaze-cueing paradigm are the best choice to investigate socio-cognitive mechanisms, due to the fact that the gaze-cueing paradigm is quite unnatural in terms of social context. Not only does it often employ 2D stimuli on the screen (which is unlike natural social interactions), but it also uses repetitive movements (i.e., the gaze-cueing paradigm employs many trials of robots gazing in one of two directions). Prior work has shown negative impacts of multiple exposures on subjective perceptions of robots over time (Bergmann et al., 2012), and recent approaches in social neuroscience (Schilbach et al., 2013) highlight the need for more ecologically valid paradigms for studying mechanisms of social cognition. Furthermore, the fact that multiple interactions over time might change our perceptions of a robot is in line with Epley et al's (2007) theory of anthropomorphism, which suggests that repetitive interactions can decrease our likelihood to attribute mental states (i.e., adopting the intentional stance) toward a robot since we no longer have the motivation to understand its behavior: effectance motivation. While prior studies have shown negative impacts of multiple exposures on subjective ratings, it is unclear if this effect can also be generalized to behavioral measurements, such as the GCEs. Specifically, it is unclear if the standard procedure of the gaze-cueing paradigm in HRI conceals aspects of social attunement, due to the fact that many trials are needed to establish an accurate and stable GCE (e.g., face-to-face studies include 160 trials; Lachat et al., 2012).

In the current study, we examined the effect of the duration of exposure to repetitive gaze behavior on both subjective perceptions of intentionality and gaze-induced attentional orienting in a gaze-cuing paradigm. Using a face-to-face gaze-cueing paradigm with a humanoid robot (i.e., iCub: Metta et al., 2008), participants were assigned to either a short exposure (i.e., 96 trials), a medium exposure (i.e., 176 trials), or a long exposure condition (i.e., 256 trials), while we collected pre- and post-ratings of adopting the intentional stance (InStance Questionnaire, Marchesi et al., 2019) and attributing mental capacities (Weisman et al., 2017). If the duration of interaction with a robot does indeed influence attribution of intentionality toward a robot, we expect that participants assigned to the different exposure conditions would experience changes in likelihood to adopt the intentional stance toward iCub after their interaction differently. To that extent, we expected to observe a different impact of interaction with the robot on the adoption of the intentional stance, depending on duration of the interaction. Additionally, we hypothesized that the relationship between adopting the intentional stance and GCEs (i.e., social attunement via attention orienting) would depend on the duration of exposure to repetitive behavior.

The aim of the current study was to examine if the duration of exposure to a humanoid robot's repetitive behavior in a gaze-cueing task influences adopting the intentional stance toward it, and if adopting of the intentional stance correlates with the GCE depending on the duration of exposure to repetitive behavior. Since previous work suggests that being exposed to robots for long durations decreases the likelihood of anthropomorphizing robots and is associated with negative affective responses, we hypothesized that participants who were exposed to more repetitive behavior would show a decrease in adopting the intentional stance in comparison to those who are exposed to less repetitive robot behavior. Additionally, we expected that the relationship between ISQ scores and GCEs would depend on exposure of the repetitive behavior of iCub (i.e., the relationship changes as a function of being exposed to repetitive gaze behavior of the robot). Finally, this study took a Bayesian approach to investigate these effects, which allows us to examine our effects in terms of probabilities and not whether an effect is significant or not.

To test our hypotheses, we employed a face-to-face gaze-cueing paradigm with iCub, and used an experimental design where observers completed a pre-questionnaire block, then the gaze-cuing experiment with iCub, then a post-questionnaire block. Results of the experiment indicated that the probability of showing a large and positive ISQ change after interacting with iCub was highest for participants who were exposed to less repetitions of iCub's gaze behavior and that the probability of showing a large and positive change decreased as a function of duration exposure. Specifically, the results showed that this probability of showing a positive ISQ difference score decreased by 73% from those who were exposed to iCub for a short duration (i.e., fewer repetitions) to those who were exposed to iCub for a medium duration. This probability decreased even further to 88% from participants who were exposed to iCub in the short exposure condition in comparison to those who were exposed to iCub for longer (i.e., long exposure condition). Interestingly, participants in the long exposure condition did not change their scores after the interaction in respect to their scores prior to interacting with iCub. This indicates that shorter interactions with a robot, even if it shows repetitive behavior, can still have a positive influence on adopting the intentional stance.

Analyses of the GCEs showed that as the length of repetition increased, GCEs were likely to become larger, as short exposure had a high probability of inducing smaller GCEs as compared to the medium exposure condition. This difference between GCEs was even more likely when comparing the short exposure condition and the long exposure condition, which indicates that more trials (i.e., repetitions) are warranted in order to measure GCEs. The results also showed that, as hypothesized, the relationship between adopting the intentional stance and gaze cueing was likely dependent on the length of exposure to repetitive robot behavior. Initially, this relationship was positive, however, as the duration of exposure increased, the relationship showed a tendency to be more negative. While this effect was not very strong, possibly due to being statistically underpowered, it suggests that longer exposures to repetitive robot behavior can reverse the positive relationship between joint attention (a marker of social attunement) and adoption of intentional stance.

The finding that short exposure durations were likely to have a positive effect on subjective ratings of adopting the intentional stance is in line with prior work that used the gaze-cueing paradigm and found that subjective ratings of the mind status of a robot increases after completing the gaze-cueing paradigm (Abubshait and Wiese, 2017; Abubshait et al., 2020b) and other studies showing that initial impressions influenced subjective ratings positively, but latter interactions kept the interactions unchanged (Paetzel et al., 2020). Furthermore, the positive impact of short interactions on adopting the intentional stance might be due to interacting with a real embodied humanoid robot, which initially might evoke social attunement, due to its physical presence (Wainer et al., 2007). However, this effect might diminish after medium and long exposure conditions, due to the repetitiveness of its behavior.

The finding that gaze cueing was likely related to adopting the intentional stance depending on exposure is supported by prior work investigating social engagement between children and social robots that shows that initial interactions with robots are paired with high social engagement but that this social engagement is reduced after a while (Komatsubara et al., 2014; Coninx et al., 2016; Ahmad et al., 2017). This suggests that the reduction in social engagement can generalize to other populations and not simply children. This account is also supported by fMRI studies that argue that the action perception network, a network in the brain that is involved in understanding and predicting others' actions (Decety and Grèzes, 1999), is activated when trying to predict movements from others' gaze behavior (Ramsey et al., 2012) and unfamiliar robotic behavior (Cross et al., 2012). After viewing repetitive actions performed by humans and robots, this activation is reduced for only robotic agents but not humans (Gazzola et al., 2007), which suggests that attention orienting to gaze cues (a phenomenon related to the social brain; Wiese et al., 2018) can be influenced by being exposed to repetitive behaviors.

The results of this study highlight an important issue in the field, which is that many of the effects of attributing intentionality to robots that we see are short-lived. Epley et al. (2007) theory of anthropomorphism provide an explanation to why, which is that when people are exposed to repetitive behaviors from an agent, it can reduce their motivation to understand the agent's behavior, which reduces the propensity to anthropomorphize the agent and adopt the intentional stance toward it. Not only do we find this short-lived effect in our data, but other studies also have found similar results. For example, when using entertainment robots, de Graaf et al. (2015) have found that the most important robot characteristic to be perceived as social is that the robot should be able to verbally respond in a free and social manner (i.e., unpredictable) and not a preprogrammed way.

This result generalizes to other areas in HRI that use repetitive interactions with a robot as interventions with populations diagnosed with autism spectrum condition (ASC). For example, Zheng et al. (2020) examined the effect of repeated exposure to a robotic platform and its effectiveness in improving the joint attention skills of children with ASC. They found that, unlike the suggestion of prior work, learned skills are not maintained over longer periods (Zheng et al., 2020). Similarly, Anzalone et al. (2019) showed a similar pattern where developing children typically did not show improvements in joint attention after prolonged exposures to a robot. These studies show that exposure over longer periods may not have a clear benefit to HRI. The results of our study shed light on this by showing that changes in subjective judgments of intentionality do change when interactions are short-lived. However, after long exposure to iCub's repetitive behavior, the change in attribution of intentionality diminished. Therefore, roboticists have to carefully investigate how longer durations of exposure can influence the human interaction partner.

The current study also illustrates how a Bayesian framework can be a better method in quantifying HRI. HRI studies often have limited sample sizes due to their complex protocols (i.e., sample sizes of about 20 participants; e.g., Mutlu et al., 2006; Terada et al., 2007; Wainer et al., 2007; Wang and Lewis, 2007; Brown and Howard, 2013; de Graaf et al., 2015; Warren et al., 2015; Zheng et al., 2020), which is problematic when analyzing data using parametric statistics. By using a Bayesian framework, we assume that the data are fixed and that the parameter estimates are dynamic, while the opposite is true for parametric methods. In other words, we assume that our parameter estimates for specific behaviors (e.g., gaze-cueing effects in reaction times) follow a distribution as opposed to it having a fixed estimate. By doing so, this method allows us to update our prior knowledge from past studies and estimate an updated parameter distribution. Not only does it allow us to update our knowledge from past studies, but it also allows roboticists to use probabilities of an effect in question. Using probabilities to quantify HRI is much more informative than using significance thresholds. For example, it is more informative for researchers to know which robot has a 70% chance of successfully engaging in joint attention with children with ASC as opposed to which robot significantly engages in joint attention with children with ASC.

Despite the advantages of the methods used here, it is also important to discuss some of the limitations associated with the current study. First, the small sample size could render the results unstable. While this could be a major issue when using parametric analyses due to violations of the general linear model's assumptions, we do not believe that the small sample size dramatically influenced our results. We believe that our results are stable due to using a Bayesian framework, and more so due to using uninformative priors (i.e., uninformative priors are considered conservative even though we could have used informed priors from prior gaze-cueing studies with iCub: e.g., Kompatsiari et al., 2018). By doing so, we inflate our credible intervals, which underestimates the confidence of our results. Another possible limitation is the use of a pre-/post-test design. In this study, participants had to complete a questionnaire that measured their attribution of intentionality to robots prior to completing the gaze-cueing experiment. This design (i.e., participation in the test prior to the experiment) could have biased the participants into attributing intentions to iCub. While this is a limitation that can be addressed by using ecologically valid behavioral measures of intention attribution (e.g., Mwangi et al., 2018), due to our controlled lab setting this method was not feasible. Still, the fact that participants did show group differences when they completed short, medium, and long gaze-cueing tasks with iCub suggests that the length of exposure does influence our interactions with robots in general.

This study provides important contribution regarding implementation of classical experimental protocols in HRI studies by showing two effects that oppose one another. First, our study showed that more trials (i.e., repetitions) were needed in order to measure a social-cognitive process that is associated with social attunement (i.e., gaze cueing), as the gaze-cueing effect was more likely more positive as repetitions increased. On the other hand, subjective ratings of adopting the intentional stance were likely less related to GCEs as repetitions (i.e., trials) increased in the gaze-cueing paradigm, which suggests that studies measuring these two constructs (i.e., social attention and adopting the intentional stance) may sacrifice the validity of one measurement over the other. In our case, the validity of social attention was likely to increase with more trials, but the validity of measuring subjective intention attribution was likely to decrease with more repetitions. Since the data suggest that we might not increase the likelihood of adopting the intentional stance toward an agent when it displays repetitive behavior over a longer duration, it poses a challenge on how to maintain sufficient number of trials for statistical power without decreasing the reception of the robot as an intentional or a human-like agent. While the gaze-cueing paradigm might be successful in evoking joint attention, it might not decrease the perception of the robot as human-like or intentional.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://osf.io/y6c9b/.

The studies involving human participants were reviewed and approved by Comitato Etico Regione Liguria. The patients/participants provided their written informed consent to participate in this study.

AA and AW conceptualized the study, interpreted the results, and wrote the manuscript. AA collected and analyzed the data. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.