Participants were recruited from the UK via Prolific. Recruitment continued until 80 valid data sets were collected for each of the three experiments; see Table 1 for demographic data. Participants received £4 for completing the study. Participants were excluded from the analysis if they reported considerable technical difficulties (e.g., stuttering of the videos, which resulted in the exclusion of n = 14 or 5% of the total number of recruited participants) or if they failed the attention check (as described in the Procedure section below, which resulted in the exclusion of n = 8 or 3% of all recruited participants).

The procedure was the same for all three experiments. They commenced with a consent form and instructions (with approval from Constructor University's Ethics Committee). The experiments then continued with the presentation of stimuli. C+ and C− stimuli were presented in alternation. The videos of one condition were consistently displayed either on the left or the right of the window, while the mapping of condition and alignment was systematically counterbalanced. This was done to ensure that participants knew which robot was presented on any given trial as their physical appearance only differed in a small white marking in the shape of either a triangle or square. Following the stimulus presentation, participants completed several questionnaires. In the first four questionnaires they rated the robots of each condition separately (participants had to answer all questionnaires first for one. then for the other robot), while the subsequent two questionnaires asked about the participants themselves. At the end of each experiment, participants were debriefed, asked what they thought the research question was and whether they experienced any technical difficulties with the study. Those participants who reported major problems (e.g., faulty stimulus presentation) were removed from the analysis (resulting in n = 14 exclusions, or 5% of all recruited participants).

The study employed a manipulation check scale (MCS) with three items for each characteristic of interest¹. E.g., for the characteristic helping the items “It appeared to be helping others,” “It appeared to understand when others needed aid” and “It seemed to seek ways in which it could help others” were used. The items were composed to range from simple descriptions devoid of selfhood-related states to descriptions referencing intentions similar to the answers of participants in Heider and Simmel's study (see the online materials at OSF for the complete questionnaire).

For accessing selfhood-attribution three already established questionnaires were used: The subscales agency and experience of the Mind-attribution Scale (MAS) (Bigman and Gray, 2018), the Godspeed Scale (GS) (Bartneck et al., 2009), as well as the Robotic Social Attributes Scale (RoSAS) (Carpinella et al., 2017). The scales were implemented as continuous scales, ranging from 0 to 100, with sliders marked only at the extremes: 0 and 100 for the MCS; “Disagree” and “Agree” for the MAS; negative and positive items for the GS and RoSAS. An attention check item asking if the robot was able to move supplemented the MAS. As the robots consistently moved in the videos, any participant who moved the response slider toward “Disagree” was excluded from the analysis (resulting in n = 8 exclusions, or 3% of all recruited participants). The MCS was completed by all participants first, and the order of MAS, GS, and RoSAS was counterbalanced with the MAS always in the middle to ensure that the attention check was placed at the same point in the study flow.

Further, we used as exploratory measures the Prior Experience with Robots Questionnaire (PERQ) (Schaefer, 2016) and the Autism Spectrum Quotient (AQ) AQ-10 (Allison et al., 2012) to gather additional information about our participants. The PERQ provides information about participants' knowledge about robots and their capabilities while the AQ indicates whether participants may be on the autism spectrum. To avoid participants being influenced by these questionnaires, they were collected after the main experiment.

All statistical analyses were conducted using R (Version 4.2.2., R Core Team, 2022). The analysis code is published in the OSF project. For each experiment, we first analyzed the data of the behavioral characteristics (manipulation check). Here ratings were aggregated by participant, cue-presence (C+ or present vs. C− or absent) and characteristic using the mean. In the analysis we conducted a two-way ANOVA with the within-participant factors of cue-presence (C+ vs. C−) and characteristic (causality, speed, equifinality, efficiency, learning sensitivity, and context sensitivity). If there was a significant interaction, we calculated post-hoc paired t-tests for cue-presence grouped by characteristic.

Next, in cases of a significant interaction between cue-presence and characteristic, we conducted a separate analysis of the selfhood-attribution data for those experiments. For this analysis we categorized in subscales rather than the overall questionnaires. The data from the MAS was therefore categorized into either agency or experience, while each rating from the GS was categorized as animacy, antropomorphism, likability, perceived intelligence, or perceived safety and RoSAS items were categorized as warmth, competence or discomfort. Ratings were then aggregated by participant, cue-presence, and subscale using the mean. This way we calculated a two-way ANOVA with cue-presence (C+ vs. C−) and subscales (agency, experience, anthropomorphism, animacy, likeability, intelligence, safety, warmth, competence, and discomfort) as within-participant factors. If there was a significant interaction, we conducted post-hoc paired t-tests for cue-presence grouped by subscale. While we included likeability and perceived safety in the analysis, we did not consider these subscales as significant contributors to accessing selfhood-attribution, rather as scales of interest in the context of social interaction or in general.

For the behavioral characteristics, the ANOVA revealed significant main effects of cue [F_{(1, 79)}= 4.12, p = 0.046, η² = 0.01] and characteristic [F_{(4.75, 375.18)}= 91.48, p < 0.001, η² = 0.30], and a significant interaction [F_{(3.93, 310.61)} = 56.38, p < 0.001, η² = 0.16]. Post-hoc paired t-tests showed significant differences for all characteristics with the C+ robot receiving higher ratings for all characteristics except speed, equifinality and efficiency (see, Figure 2A left).

The analysis of the selfhood-attribution data also showed significant main effects of cue-presence [F_{(1, 79)} = 10.45, p = 0.002, η² = 0.02] and subscale [F_{(4.08, 322.29)} = 171.38, p < 0.001, η² = 0.45], and a significant interaction [F_{(4.92, 388.54)} = 25.68, p < 0.001, η² = 0.07]. Post-hoc paired t-tests further revealed significant differences for all subscales except agency [t(79) = 0.05, p = 0.958], anthropomorphism [t(79) = 1.68, p = 0.098] and competence [t(79) = 0.08, p = 0.937] (all other p-values <= 0.007). Ratings were higher for the C− robot on all significant subscales except intelligence (see, Figure 2A right).

For the behavioral characteristics, we observed a significant main effect of cue-presence [F_{(1, 79)} = 20.36, p < 0.001, η² = 0.03] and characteristic [F_{(5.27, 416.65)} = 55.00, p < 0.001, η² = 0.20], and a significant interaction [F_{(5, 395.31)} = 22.37, p < 0.001, η² = 0.04]. Post-hoc paired t-tests showed significant differences for the characteristics learning [t(79) = 4.07, p < 0.001] and context sensitivity [t(79) = 2.52, p = 0.014], attention sharing [t(79) = 7.73, p < 0.001] and helping [t(79) = 4.08, p < 0.001]. The C+ robot was rated higher than the C− robot on these characteristics (see, Figure 2B left).

Next, the analysis of the selfhood-attribution data showed significant main effects of cue-presence [F_{(1, 79)} = 27.98, p < 0.001, η² = 0.03] and subscale [F_{(4.35, 343.76)} = 130.18, p < 0.001, η² = 0.39], and a significant interaction [F_{(3.82, 301.84)} = 9.04, p < 0.001, η² = 0.02]. Post-hoc tests revealed significant differences in all subscales except agency [t(79) = 1.36, p = 0.178] and perceived safety [t(79) = 1.23, p = 0.221] (all other p-values <= 0.006). The ratings for the C+ robot were higher on all subscales except for discomfort (see, Figure 2B right).

First, for the behavioral characteristics we observed significant main effects of both cue-presence [F_{(1, 79)} = 31.65, p < 0.001, η² = 0.04] and characteristic [F_{(4.27, 337.51)} = 55.51, p < 0.001, η² = 0.24], and a significant interaction [F_{(3.97, 313.56)} = 20.57, p < 0.001, η² = 0.04]. Almost all post-hoc paired t-tests were significant except for speed [t(79) = 0.18, p = 0.860], behavioral efficiency [t(79) = 0.59, p = 0.554] and learning sensitivity [t(79) = 1.61, p = 0.112] (all other p-values <= 0.050). On the significant comparisons the C+ robot was always rated higher than the C− robot (see, Figure 2C left).

Next, the ANOVA of the selfhood-attribution revealed significant main effects of cue-presence [F_{(1, 79)} = 51.01, p < 0.001, η² = 0.08] and subscales [F_{(3.90, 307.75)} = 165.53, p < 0.001, η² = 0.51], and a significant interaction [F_{(3.51, 277.49)} = 18.11, p < 0.001, η² = 0.04]. Post-hoc tests showed significant differences on all of the tested subscales (all p-values <= 0.003) with ratings for the C+ robot always higher than the C− robot, except for the subscale discomfort—where the C− robot (M = 23.99, SD = 25.42) received higher ratings than the C+ robot (M = 17.22, SD = 20.72) (see, Figure 2C right).

To identify commonalities across the 3 experiments, we calculated Pearson's correlation coefficients for ratings of the characteristics and selfhood-attribution subscales across all experiments. This analysis showed that a majority of characteristics and subscales have significant, positive correlations (63% of all correlation-coefficients), with moderate coefficients accounting for 23% and strong coefficients accounting for 14% of all coefficients (see, Figure 3 for a visual representation).

Note that for the subscale discomfort almost all of the significant correlations are negative, which makes sense as this is the only negatively framed subscale. This subscale further shows an over proportionate number of non-significant correlations suggesting that it is not a crucial dimension for selfhood-attribution.

Next, to characterize the results of the agency attribution we calculated Bayes factors testing for an effect of cue presence on agency using the r package “BayesFactor” (Morey and Rouder, 2024) and following interpretations guidelines as outlined by Kass and Raftery (1995). This was done for each experiment presented here and for the experiments presented in our previous paper (Pohl et al., 2024, under review).

For the social experiments, this analysis reveals decisive evidence for an effect of cue-presence on agency for the last experiment (helping, BF₁₀ > 100) and for the first two experiments no evidence for an effect of agency (social sensitivity: BF₁₀ = 0.05; attention sharing: BF₁₀ = 0.44) and in the case of social sensitivity strong evidence against such an effect (BF₀₁ = 19.41).

For the non-social experiments, we found decisive evidence for an effect of cue-presence on agency for the experiments manipulating causality (BF₁₀ > 100), equifinality (BF₁₀ > 100) and learning sensitivity (BF₁₀ > 100) as well as strong evidence for the experiment manipulating efficiency (BF₁₀ = 40.70). Surprisingly, even for the last experiment (context sensitivity), where the manipulation check did not reveal any significant differences in the perception of the characteristics, we found substantial evidence (BF₁₀ = 9.19) for an effect of cue-presence on agency.

Since the situations presented in our experiments were social in nature we additionally investigated whether participants' attribution toward the robots was modulated by autism. For this purposes we calculated a three-way ANOVA with within-subjects factors cue-presence and subscale, as well as the between-subject factor Autism Spectrum Quotient (AQ). The ANOVA revealed a significant main effect of the AQ [F_{(1, 25, 400)} = 99.93, p = 0.004], with overall lower ratings with higher scores (see, Figure 4). Further there were significant two-way interactions of the AQ both with subscale [F_{(9, 25, 400)} = 2.73, p = 0.003] and with cue-presence [F_{(1, 25, 400)} = 7.01, p = 0.008], but no significant three-way interaction [F_{(9, 25, 400)} = 1.53, p = 0.132]. A descriptive analysis of the data suggests that participants with a higher AQ are more sensitive to the manipulation with a stronger decrease in ratings for the C− robots as compared to the C+ robot (see, Figure 4).

An exploration on the effect of prior experience with robots on participant's perception of the robots was done by Nikolovska and colleagues (2024), suggesting that participants more experienced with robots rate the social C+ robot more favorable as compared to the non-social C− robot.

Two outcomes of the first experiment are noteworthy. First, in line with our previous results, participants over-generalized self-related ratings beyond the manipulated characteristic; they rated the critical robot higher on all characteristics except speed, equifinality, and efficiency, indicating that they perceived the C+ robot as exhibiting causal behavior, learning, context and social sensitivity, as well as exhibiting attention sharing and helping behavior. The control robot, in turn, was perceived as moving with more human-like speed and as exhibiting more goal-directed and efficient behavior. Second, participants rated the critical robot higher on most of the selfhood-attribution scales, although this did not include agency. In general, the data support our previous finding that the presence of a single behavioral cue results in the greater attribution of selfhood to the agent exhibiting a critical cue in addition to activating representations of characteristics that are missing behavioral cues in a pattern similar to what has been described as the Halo effect (e.g., Forgas and Laham, 2016).

In the second experiment, participants again rated the critical robot higher on more than one behavioral characteristic, though this time this was limited to learning and context sensitivity for the non-social characteristics, and attention sharing and helping for the social ones. While the generalization of characteristics was more limited than in Experiment 1, it still extended beyond the manipulated characteristic. Regarding the selfhood-attribution the C+ robot was rated significantly higher on all subscales except agency and safety. Hence, we replicated the observation from the first experiment that critical and control robot were not perceived to differ in their agency. While the literature suggests the importance of social factors in the attribution of agency (e.g., Carrier et al., 2014; Khamitov et al., 2016; van der Woerdt and Haselager, 2019) there are barely any studies directly investigating the effect of mere sociality. However, there is some literature on the complex link between the perception of sociality and warmth to the attribution of agency, which seems to be modulated by valence (see Suitner and Maass, 2008).

In the third experiment, participants again rated the critical robot higher on a majority of the (non-manipulated) characteristics including all social characteristics (social sensitivity, attention sharing and helping) as well as causality, equifinality, and context sensitivity. Regarding the selfhood-attribution, the C+ robot was also rated higher on all subscales except discomfort—which is the only negatively poled subscale. Overall, this indicates that participants perceived the critical robot not only as more helping, but also as more causal, goal-directed and context sensitive, as well as more socially sensitive and exhibiting more attention sharing. Participants further had the stronger tendency to attribute a self to the critical robot, unlike in Experiment 1 and 2 this effect was also observed for the subscale agency.

Across experiments, the exploratory analysis shows that, in line with the general literature (e.g., Castelli et al., 2002; Klin et al., 2003), participants with a higher indication for autism attribute less selfhood to the robots (i.e., they show a stronger tendency to take them at face-value). Interestingly, individuals with an indication for autism exhibited a stronger difference between the two conditions, implicating that they were more sensitive to the presence of social behavior than individuals with low indication for autism. Moreover, the additional analysis of the impact of people's prior experience with robots done by Nikolovska and colleagues (2024) reveals that people who are more knowledgeable about robots also were more sensitive to the manipulation of sociality in the robots. Overall, this evidence is in favor of our assumption that in attributing selfhood to other agents people refer to a subjective concept which does differ between individuals.

Interestingly for our purposes, the impact of manipulating these social characteristics was not restricted to the corresponding scale. More specifically, our manipulations also induced two other kinds of effect. First, and particularly important for our present purposes, the higher scores for the manipulated characteristics over-generalized to the other social scales. That is, agents showing socially sensitive behavior are also assumed to share attention with or help others, even in the absence of any perceptual evidence of such behavior, and the same kind of over-generalization was found for manipulations of social attention and helping. In other words, agents that are seen to engage in one kind of social behavior are also assumed to engage in other kinds of social behavior, even in the absence of any supporting data. This might be taken to reflect the attribution of some kind of social “personality,” of which the observed social behavior is only one example.

Second, however, the observed over-generalization was not restricted to other social characteristics. Interestingly, this kind of over-generalization could go two ways. On the one hand, some of the non-social scales showed higher ratings for the robot engaging in social activities. For instance, the socially sensitive C+ in the first experiment was also considered to show more causal behavior, learning, and context sensitivity, and similar patterns were obtained for the agents showing social attention and helping. On the other hand, however, engaging in social behavior also led to a perceived decrease with respect to other non-social characteristics. For instance, the socially sensitive robot in Experiment 1 was perceived as less oriented on the same goal (equifinality) than the control robot, which suggests some kind of trade-off between individual goal pursuit and social responsiveness.

Indeed, the way we manipulated social sensitivity (e.g., by having the critical robot stop its own navigation and turning to the other robot) reflects this trade-off. While one may consider that an experimental flaw, we take it as a valid reflection of a conceptual trade-off, in the sense that the sociality of one's behavior is defined by the degree to which one is willing to neglect one's own goal pursuit to the benefit of others. Accordingly, we consider the negative effects of engaging in social behavior on scales assessing non-social characteristics as indications of a conceptual relationship between sociality and individualism, and the respective concepts. The perhaps strongest indication of this kind of individual-social trade-off relates to agency. Whereas our previous manipulations of non-social selfhood characteristics almost always had a very strong (positive) impact on agency, this agency advantage was absent in two of our three experiments. Hence, the manipulation of social behavior has a much lesser impact on the attribution of agency than the manipulation of non-social characteristics. A similar trade-off can be found in the phenomenon described as Bystander effect (e.g., Fischer et al., 2011), which shows that being or merely imagining to be in a social situation reduces people's willingness to help (Garcia et al., 2002). The reasoning here is, that being in a group creates a diffusion of responsibility—which has been linked to a reduced sense of agency (Beyer et al., 2017). We hypothesize that this first hand experience of a disconnect between one's own actions and their outcomes, when being in a group, may lead to also perceiving other agents as having less individual agency given that they are part of a group.

Observing that people differ in how they attribute selfhood to other agents is in line with our assumption of naïve concepts of selfhood, that is, we investigated what individuals' idea of what the self is, rather than an objective definition. Considering that even in the scientific debate there are many different concepts of selfhood, it is even less surprising that in everyday life different people have different notions of what the self is. We provide evidence for a systematic link between how this selfhood-concept is attributed and, both, dispositional factors such as autism as well as more situational factors such as familiarity with the perceived agent.

Overall, the naïve approach to selfhood-attribution and the Pars-Pro-Toto account offer a novel perspective on investigating selfhood. While many theories of mind perception commonly either have a narrow focus [e.g., on agency and experience, Gray et al. (2007)] or are concerned with inferring specific mental states such as intentions (e.g., Theory of Mind, Frith and Frith, 2005; and Intentional Stance, Dennett, 1989) the Pars-Pro-Toto account takes a naïve approach that takes into account that people may differ in their ideas of what the self is. Thus, we bridge concepts like the “mind,” agency, consciousness, higher cognitive functions and sociality by considering selfhood as an umbrella term (as shown in Figure 1). Critically, our studies show, that people tend to start attributing their concept of selfhood when they perceive an agent exhibiting behavioral cues for characteristics deemed relevant to selfhood in the literature. Moreover, people seem to frequently over-generalize from cues for any single characteristic to others. Minimal information, thus, seems sufficient to activate the entire selfhood concept, even when other cues are absent. While this is reminiscent of the Halo-effect, our account describes an internal conceptual over-generalization rather than an evaluative bias reflecting correlational misattribution across traits.

Our findings stress the general tendency of humans to anthropomorphize agentic systems. While there is not a large body of literature investigating the movement of robots, a few examples suggest that even robot arms when moving in a human-like manner elicit anthropomorphization (e.g., Hostettler et al., 2023; Kupferberg et al., 2012). Our study extends on these findings suggesting that people even anthropomorphize small vehicle-like robots when they move in way that suggests social characteristics like sensitivity to others, attention sharing, and helping behavior; or as shown in our previous study (Pohl et al., 2024, under review) also for non-social characteristics like efficiency or learning sensitivity. Thus, anthropomorphism seems less a product of technological sophistication than of human cognitive tendencies. A possible explanation given by Urquiza-Haas and Kotrschal's (2015) is a motor matching mechanism inspired by (embodied) simulation theory (Goldman, 2006; Svensson et al., 2008). They argue that the mirror neuron system responsible for simulating others' actions is less dependent on the physical body of the observed agent but sensitive to actions familiar to the observer (see Buccino et al., 2004). Following, a robot of any shape navigating the physical world may trigger an observer's simulation of doing the same, thus driving anthropomorphization.

Considering the advance of technological systems into society the question is raised about moral status of artificial agents. Robots may be perceived as moral agents based on perceived rather than actual capacities, leading to misplaced moral consideration for machines and potentially creating conflicts when interacting in the same physical space (see, e.g., the perceptual believe problem, Thellman and Ziemke, 2021). This highlights the need for technological systems, especially embodied systems found in human spaces, to be designed carefully with “the right amount of anthropomorphism.” On the one side it is desirable that human-likeness is attributed, as this drives trust (e.g., Natarajan and Gombolay, 2020), on the other side, it may foster misconceptions about a system's moral status and cognitive abilities. Moreover, if selfhood attributions emerge from an observer's social expectations rather than ontological realities, then debates on robot rights or moral status may need to shift from asking whether machines possess a self to how human practices of attribution shape the moral and legal landscapes².

Taken together, our study sheds light on the contribution of social factors to attributing selfhood to other agents. We showed that simple social and behavioral cues related to social sensitivity, attention sharing, and helping are relevant for attributing selfhood. We replicated our previous finding that people tend to over-generalize from one manipulated cue to other, non-manipulated cues. However, this over-generalization was more systematic within the social domain, in the sense that people tend to generalize from observed to unobserved social behavior, whereas the generalization to non-social characteristics sometimes follows other rules. With respect to some cues, this over-generalization works as within the social domain, in the sense that engaging in social behavior is assumed to imply other non-social characteristics. With respect to other cues, however, this relationship is negative, in that engaging more in social behavior reduces the perception of non-social characteristics. Further, we observed that people did not consistently attribute more agency to the social robot which may reflect the conceptual basis of how we construe sociality vis-à-vis individual goal pursuit.