Do We Adopt the Intentional Stance Toward Humanoid Robots?

In daily social interactions, we need to be able to navigate efficiently through our social environment. According to Dennett (1971), explaining and predicting others’ behavior with reference to mental states (adopting the intentional stance) allows efficient social interaction. Today we also routinely interact with artificial agents: from Apple’s Siri to GPS navigation systems. In the near future, we might start casually interacting with robots. This paper addresses the question of whether adopting the intentional stance can also occur with respect to artificial agents. We propose a new tool to explore if people adopt the intentional stance toward an artificial agent (humanoid robot). The tool consists in a questionnaire that probes participants’ stance by requiring them to choose the likelihood of an explanation (mentalistic vs. mechanistic) of a behavior of a robot iCub depicted in a naturalistic scenario (a sequence of photographs). The results of the first study conducted with this questionnaire showed that although the explanations were somewhat biased toward the mechanistic stance, a substantial number of mentalistic explanations were also given. This suggests that it is possible to induce adoption of the intentional stance toward artificial agents, at least in some contexts.


Sample
First, we collected data of one hundred and twenty Italian native speakers with different social and educational backgrounds (see Table 1 for demographical details) who completed our Human Control Questionnaire (HCQ): HCQ mapping not-counterbalanced, N=120). Due to a mistake in counterbalancing the position (left vs. right) of mechanistic or mentalistic descriptions, we administered the HCQ with the proper counterbalanced mapping to a second sample of hundred and three participants, from which we excluded one participant who already filled in the first HCQ: HCQ mapping counterbalanced, N=102. Data collection was conducted in accordance with the ethical standards laid down in the Code of Ethics of the World Medical Association (Declaration of Helsinki), procedures were approved by the regional ethics committee (Comitato Etico Regione Liguria). Table 1. Demographic details of the sample (N=222).

Demographic Characteristic
Age (years), mean (SD) [min, max] 29.81 (10.39) [18,68] Female, n (%) 153 (68.92) Education (years), mean (SD) [min, max] 16.20 (2.97) [8,24] Human Control Questionnaire (HCQ) We selected 15 scenarios out of the 34 original ones that could be adapted to a human agent with respect to the mechanistic descriptions (we excluded all items that used very implausible mechanistic descriptions such as motor calibration) and digitally edited them (Adobe Photoshop CC 2018) depicting a human agent (Paola) instead of the iCub robot. Each item of the HCQ was identical with the ISQ, except for the depicted agent in the scenarios, which for HCQ was a human (see Figure 1 for an example, but see also Supplementary Material 3 for all the scenarios included in the HCQ). Each scenario was composed of three pictures (size 900 x 173.2 pixels). Out of the 15 scenarios, 9 involved one (or more) other human interacting with Paola; 1 scenario showed a human arm pointing to an object; 6 scenarios depicted only Paola.
The types of action performed by Paola in the scenarios were: grasping, pointing, gazing, and head movements.
As in the original ISQ, each item included two sentences, in addition to the scenario. One of the sentences was always explaining Paola's behaviour referring to the design stance (i.e., mechanistic explanation), whereas the other was always describing Paola's behaviour referring to mental states (i.e., mentalistic explanation). We kept the human agent's emotional expression constant within and across the scenarios not to bias towards mentalistic explanations, see Supplementary Material 3 for a complete list of items.

Data Analysis and Results.
All statistical analyses were performed in R (version 3.4.0, freely available at http://www.rproject.org). Data analysis was conducted in three steps. First, we analysed responses in the HCQ as we did for the InStance questionnaire (ISQ). Second, we selected from our original ISQ only responses to the same 15 items included in the HCQ, and we performed analyses on those. Finally, we compared responses in the two questionnaires (15 items) between groups.

Human control questionnaire (HCQ)
For each participant, we calculated the average score. As we did for the ISQ, we converted the bipolar scale into a 0-100 scale where 0 corresponded to completely mechanistic and 100 to a completely mentalistic explanation. The null value of the scale, i.e. the starting position of the slider that was equally distant from both limits, corresponded to 50. Scores under 50 meant the answer was 'mechanistic', scores above 50 meant they were 'mentalistic'. Firstly, we investigated the effect of the mapping by comparing the average score across groups (i.e. notcounterbalanced mapping sample vs. counterbalanced mapping sample). Independent sample t-test showed that the average score did not differ across groups, t (220) < 1. Given that no differences were found between mappings, from the total sample of N=222 participants, we randomly selected 106 respondents, to match the sample size of our original ISQ. The overall average score for the HCQ was 54.62 (with 0 value indicating the most mechanistic score and 100 indicating the most mentalistic score). We tested the distribution of the average score for normality with the Shapiro-Wilk test. Results showed that the average scores were distributed normally, W = 0.99, p > .05. Moreover, in order to check that the average score for the HCQ was not the result of random choice, we conducted one-sample t-tests against a critical value of 50 (i.e. the position at which the slider was equally distant from both statements). Results showed that the average score significantly differed from 50, t (105) = 3.67, p < .001). Then, as we did for the ISQ, we focused only on respondents who were not familiar with robots. From the original sample of N=222, we selected randomly N= 89 respondents (to match the sample of the non-familiar group in ISQ), who reported no familiarity with robots. The average score was 54.82, and it was distributed normally, Shapiro-Wilk test: W = 0.97, p = .052. Then, as we did for the ISQ, we estimated the percentage of participants who attributed 'mechanistic' or 'mentalistic' descriptions according to their average score. Participants who scored below 50 (0 -50 in our scale) were assigned to the Mechanistic group (N=32), whereas participants with an average score above 50 (50 -100)

Comparison between InStance and Human control questionnaire
Average scores for ISQ and HCQ for 15 items are shown in