1
2
3
4 Do we adopt the Intentional Stance toward humanoid robots?
5
6
7 Serena Marchesi1, Davide Ghiglino1,2, Francesca Ciardo1, Ebru Baykara1, Agnieszka Wykowska1*
8 1Istituto Italiano di Tecnologia, Genoa, Italy
9 2DIBRIS, Università di Genova
10 11 *Corresponding author: 12
13 Agnieszka Wykowska 14 Istituto Italiano di Tecnologia 15 Centre for Human Technologies 16 Via E. Melen, 83 17 16040 Genova, Italy
18 e-mail: [email protected] 19 20
1
21 Abstract
22 In daily social interactions, we explain and predict behaviours of other humans by referring to mental states
23 that we assume underlie their behaviours. In other words, we adopt the intentional stance (Dennett, 1971)
24 towards other humans. However, today we also routinely interact with artificial agents: from Apple’s Siri
25 to GPS navigation systems. In the near future, we will casually interact with robots. Since we consider
26 artificial agents to have no mental states, we might not adopt the intentional stance towards them, and this
27 might result in not attuning socially with them. This paper addresses the question of whether adopting the
28 intentional stance towards artificial agents is possible. We propose a new method to examine if people have
29 adopted the intentional stance towards an artificial agent (humanoid robot). The method consists in a
30 questionnaire that probes participants’ stance by requiring them to choose the likelihood of an explanation
31 (mentalistic vs. mechanistic) of behaviour of a robot depicted in a naturalistic scenario (a sequence of
32 photographs). The results of the first study conducted with this questionnaire showed that although the
33 explanations were somewhat biased towards the mechanistic stance, a substantial number of mentalistic
34 explanations was also given. This suggests that it is possible to induce adoption of the intentional stance
35 towards artificial agents. Addressing this question is important not only for theoretical purposes of
36 understanding the mechanisms of human social cognition, but also for practical endeavours aiming at
37 designing social robots.
38
39
40
2
41 Introduction
42 Over the last decades, new technologies have entered inexorably our houses, becoming an integral part of
43 our everyday life. Our constant exposure to digital devices, some of which seemingly “smart”, makes the
44 interaction with technology increasingly more smooth and dynamic, from generation to generation (Baack,
45 Brown &Brown, 1991; Zickuhr, K., & Smith, A., 2012; Gonzàles, Ramirez, Viadel, 2012). Some studies
46 support the hypothesis that this exposure is only at its beginning: it seems likely that technologically
47 sophisticated artifacts, such as humanoid robots, will soon be present in our private lives, as assistive
48 technologies and housework helpers (for a review see Stone et al., 2017).
49 Despite the technological habituation, we are (and presumably we will continue to be) undergoing, little is
50 known about social cognition processes we put in place during interaction with machines, humanoid robots
51 specifically. Several authors have theorized that humans possess a natural tendency to anthropomorphize
52 what they do not fully understand. Epley et al. (2007) for instance, defined anthropomorphism as the
53 attribution of human-like characteristics and properties to non-human agents and/or objects, independent
54 of whether they are imaginary or real. The likelihood of spontaneous attribution of anthropomorphic
55 characteristics depends on three major conditions (Epley et al., 2007; Waytz, et al, 2010): first, availability
56 of characteristics that activate existing knowledge that we have about humans ; second, the need of social
57 connection and efficient interaction in the environment; and finally individual traits (such as need of
58 control) or circumstances (e.g., loneliness, lack of bonds with other humans). In the vol. I of the Dictionary
59 of the History of Ideas, Agassi argues that anthropomorphism is a form of parochialism allowing projecting
60 our limited knowledge into a world that we do not fully understand. Some other authors claimed that we,
61 as humans, are the foremost experts in what it means to be human, but we have no phenomenological
62 knowledge about what it means to be a non-human (Nagel, 1974; Gould, 1996). For this reason, when we
63 interact with entities for which we lack specific knowledge, we commonly choose “human” models to
64 predict their behaviors.
65 This stance might be taken also when we interact with technology, the operation of which we cannot
66 understand. Almost everyone has seen someone shouting at their malfunctioning printer, or imploring their
3
67 phone to work, as if they were interacting with another person. Thinking in rational terms, no adult in the
68 healthy population would ever consider his/her smartphone as a sentient being, yet, sometimes, we behave
69 as if technological devices around us possessed intentions, desires and beliefs.
70 This has an important theoretical implication: from this perspective, it seems that other’s mental states are
71 a matter of external attribution rather than a propriety of an entity. Waytz et al. (2010) support this
72 hypothesis, and consider the attribution of mental states as the result of perception of an observer of an
73 observed entity, as a sum of motivational, cognitive, physical and situational factors. Other authors found
74 that several cues can foster the attribution of mental states to non-human entities, from motion (Heider,
75 Simmel, 1944; Scholl, Tremoulet, 2000), and physical properties of an object (Morewedge et al., 2007) to
76 the personological characteristic of the observer (Epley et al., 2007). Furthermore, it seems that the
77 tendency to attribute mental states is ubiquitous and unavoidable for human beings (Parlangeli et al, 2012).
78 Anthropomorphizing seems to be a default mode that we use to interpret and interact with the agents with
79 which we are sharing (or going to share) our environment (Epley et al., 2007, Waytz et al., 2010, Wiese et
80 al., 2017), unless other manners of interpreting other entities’ behavior are available; for example, if a
81 coffee machine is not working because it is not connected to a power socket, we would certainly not explain
82 the failure in delivering the coffee by the coffee machine with reference to its bad intentions.
83 Already in the seventies, Dennett proposed that humans use different strategies to explain and predict other
84 entities’ (objects, artifacts or conspecifics) behaviors (Dennett, 1971, 1987). Dennett defines three main
85 strategies or “stances” that humans use. Consider chemists or physicists in their laboratory, studying a
86 certain kind of molecules. They try to explain (or predict) the molecules’ behavior through the laws of
87 physics. This is what Dennett calls physical instance. There are cases in which laws of physics are an
88 inadequate (or not the most efficient) way to predict the behavior of a system. For example, when we drive
89 a car, we can fairly predict that the speed will decrease if we push the brake pedal, since the car itself is
90 designed this way. To make this kind of prediction, we do not need to know the precise physical
91 mechanisms behind all atoms and molecules in the braking system of the car, but it is sufficient to rely on
92 our experience and knowledge of how the car is designed. This is described by Dennett as the design stance.
93 Dennett proposes the existence of also a third strategy, the intentional stance. Intentional stance relies on
4
94 ascription of beliefs, desires, intentions and, more broadly, mental states to a system. Dennett (1987)
95 defines each system whose behaviour is predictable by others with reference to mental states as an
96 intentional system. Epistemologically speaking, Dennett refers to mental states in a realistic manner: he
97 treats mental states as an objective phenomenon. This is because adopting the intentional stance is the best
98 strategy only towards truly intentional systems (the “true believers”), while for other systems, other
99 strategies might work better. What makes Dennett an interpretationist, however, is that according to his
100 account, mental states can be discerned only from the point of view of the one who adopts intentional stance
101 as a predictive strategy, and dependent on whether this strategy is successful or not. In other words, despite
102 beliefs being a real phenomenon, the success of a predictive strategy confirms their existence.
103 Interestingly, adopting the intentional stance towards an artefact (such as a humanoid robot) does not
104 necessarily require that the artefact itself possesses true intentionality. However, adopting the intentional
105 stance might be a useful or default (as argued above) way to explain a robot’s behavior. Perhaps we do not
106 attribute mental states to robots but we treat them as if they had mental states. Breazeal (1999) highlighted
107 that this process does not require endowing machines with mental states in the human sense, but the user
108 might be able to intuitively and reliably explain and predict the robot’s behavior in these terms. Intentional
109 stance is a powerful tool to interpret other agents’ behavior. It leads to interpret behavioral patterns in a
110 general and flexible way. Specifically, flexibility in changing predictions about others’ intentions is a
111 pivotal characteristic of humans. Adopting the intentional stance is effortless for humans, but of course, it
112 is not the perfect strategy: if we realize that this is not the best stance to make predictions, we can refer to
113 the design stance or even the physical stance. The choice of which stance to adopt is totally free and might
114 be context-dependent: it is a matter of which explanation works best. Let us consider our interactions with
115 smartphones. The way we approach them is fluid: we adopt the design stance when we hear a “beep” that
116 notifies us about a text message or an e-mail, but we might switch to the intentional stance when a voice
117 recognition such as Apple’s Siri is not responding to us adequately, and gives us an incorrect answer to our
118 question. In fact, it might even happen that we become frustrated and ask our smartphone a rhetorical
119 question “why are you so stupid?”. In this context, it is also important to consider cultural differences: the
120 likelihood of adopting the intentional stance towards artefacts might differ from culture to culture (Dennett,
5
121 1987). Learning how humans approach non-human agents and explain their behavior (with reference to
122 intentional or design stance) is crucial for philosophical and scientific research and also for social robotics.
123 This is because adopting intentional stance might facilitate treatment of artefacts more socially. This, in
124 turn, might ease introduction of these new technologies in our daily lives. Although these forms of
125 “synthetic intelligence” are considered as future assistive technologies, with potential applications from
126 healthcare to industry, some studies brought to light potential issues with acceptance of artefacts in the
127 human social environments (Bartneck and Reichenbach, 2005). Furthermore, pop-culture has induced a
128 substantial amount of skepticism towards robots, associating them with threats for humanity (Kaplan,
129 2004). For these reasons, it is crucial to understand what factors contribute to acceptance of robots in human
130 environments, and whether adoption of intentional stance is one of those factors. It has been argued that
131 robots, and specifically humanoids, have the potential to trigger mind attribution, as long as they display
132 observable signs of intentionality, such as human-like behaviors (Wykowska et al., 2015 a,b) and
133 appearance (i.e. biological motion or human–like facial traits) (Wiese, 2017, Chaminade et al., 2007).
134 Currently, however, most experimental protocols addressing the question of adoption of intentional stance
135 towards artificial agents rely on the manipulation of attribution of intentionality via instruction (Gallagher
136 et al., 2002; Chaminade et al., 2012; Wiese, Wykowska et al., 2012; Wykowska, Wiese et al, 2014). To our
137 knowledge, no previous studies systematically and empirically explored spontaneous adoption of
138 intentional stance toward artificial agents. In this context, we created a tool that should allow examining
139 whether people adopt the intentional stance towards a humanoid robot (the iCub robot, Metta et al., 2008).
140 We created thirty-five fictional scenarios, in which iCub appeared (in a series of photographs) engaged in
141 different activities in a daily life context. Each scenario consisted of three aligned pictures, depicting a
142 sequence of events. For each scenario, participants had to rate (by moving a slider on a scale) if iCub’s
143 behavior is motivated by a mechanical cause (referring to the design stance, such as malfunctioning,
144 calibration, etc.) or by a mentalistic reason (referring to the intentional stance, such as desire, curiosity,
145 etc.).
6
146 Method and Materials
147 Sample
148 One hundred and six Italian native speakers with different social and educational backgrounds (see Table
149 1 for demographical details) completed our InStance questionnaire. Data collection was conducted in
150 accordance with the ethical standards laid down in the Code of Ethics of the World Medical Association
151 (Declaration of Helsinki), procedures were approved by the regional ethics committee (Comitato Etico
152 Regione Liguria).
153 Table 1. Demographic details of the sample (N=106).
Demographic Characteristic
Age (years), mean (SD) [min, max] 33.28 (12.92) [18,72]
Female, n (%) 68 (64.2)
Education (years), mean (SD) [min, max] 16.43 (3.04) [8, 24]
154
155 Questionnaires
156 InStance
157 Each item of the InStance questionnaire was composed of a sequence of three images and two sentences
158 with a scale and a slider between the two sentences (one of the sentences was positioned on left, and the
159 other one on the right extreme of the scale), see Figure 1 for an example. For a complete list of images and
160 sentences included in the questionnaire see the supplementary material (English version), or the following
161 link (for English and Italian version): https://instanceproject.eu/publications/rep.
7
162
163 Figure 1. Screenshot from the InStance Questionnaire in English.
164 We created 35 scenarios depicting the iCub robot interacting with objects and/or humans. Each scenario
165 was composed of three pictures (size 800x173.2 pixels). Out of the 35 scenarios, 13 involved one (or more)
166 human interacting with the robot; 1 scenario showed a human arm pointing an object; 22 scenarios depicted
167 only the iCub robot. 10 scenarios included pictures digitally edited (Adobe Photoshop CC 2018). The types
168 of action performed by iCub depicted in the scenarios were: grasping, pointing, gazing, and head
169 movements. All persons depicted in the photographs provided written consent to be depicted in the
170 photographs used in our questionnaire.
171 Each item included two sentences, in addition to the series of three pictures. One of the sentence was always
172 explaining iCub’s behaviour referring to the design stance (i.e., mechanistic explanation), whereas the other
173 was always describing iCub’s behaviour referring to mental states (i.e., mentalistic explanation).
174 In order to be certain that the sentences were describing the action in a mechanistic or mentalistic way,
175 prior to data acquisition, we distributed the sentences (only sentences alone, with no associated pictures)
176 to 14 volunteers who had a degree in philosophy (all Italian native speakers) to rate on a 10-points Likert
177 scale how much they understood each sentence as mechanistic or mentalistic (0 = totally mechanistic, 10
8
178 = totally mentalistic). As no scenario was presented with the sentences, the raters were not informed that
179 the sentences were created to describe the behaviour of a humanoid robot. In addition, the subject of each
180 sentence was named as a general “Agent A” to avoid any bias arising from the name of the robot.
181 Based on the responses in this survey, we modified the sentences that were not matching our intent. In
182 particular, we modified sentences that obtained an average score between 4 and 5 (meaning that they were
183 not clearly evaluated as mechanistic or mentalistic). We modified 12 out of the 35 initial pairs of sentences
184 to match our intended description (mentalistic or mechanistic).
185 Other questionnaires
186 In addition to the InStance questionnaire, we administered the Italian version of the Reading the Mind in
187 the Eyes (Baron-Cohen et al., 1999, Baron-Cohen et al., 2001, Serafin & Surian, 2004) and the Theory of
188 Mind subscale of the questionnaire developed by Völlm (2006). These tests were used as control to check
189 for outliers in Theory of Mind abilities.
190 Reading the Mind in the Eyes
191 The Reading the Mind in the Eyes test was developed by Simon Baron-Cohen and colleagues (1999, 2001)
192 to test the abilities of inferring other’s mental states through only looking at others’ eyes. In this
193 questionnaire, participants are asked to view 36 photos of people’s eyes. Below the photographs of the
194 eyes, four adjectives are presented. Participants are asked to choose the one adjective that describes best
195 the photograph they are looking at. We selected this test as it is one of the most used tests to measure
196 Theory of Mind abilities.
197 A test of Völlm and colleagues (2006)
198 Völlm et al. (2006) developed a questionnaire to evaluate neural correlates of Theory of Mind and empathy
199 in an fMRI study. In this test, the presented stimuli are non-verbal cartoon stripes of three frames:
200 participants are instructed to look at the stripes and choose the frame that, according to them, would best
201 conclude the depicted story. For the purpose of this study, we presented only the 10 items included in the
9
202 Theory of Mind subscale of the test. We selected this test as a non-verbal test of mentalising abilities,
203 complementary to the Reading the Mind in the Eyes test.
204 Data acquisition for the InStance questionnaire
205 All the questionnaires were administered via SoScisurvey (www.soscisurvey.de). Participants received the
206 URL addresses of all the questionnaires. They gave informed consent online to participation in the survey
207 by clicking “I agree” button on the consent form site under the provided links, before the questionnaires
208 began. Participants were asked to fill out the questionnaires in the order they were provided: first the
209 InStance, then the Mind in the Eyes and finally the Völlm questionnaire. The InStance Questionnaire was
210 composed by 35 items and 1 example item. Only one item at a time was presented (cf. Figure 1).
211 In each item, participants were explicitly instructed to move a cursor on a slider scale toward the sentence
212 that, in their opinion, was a more plausible description of the story depicted in the scenario. The two
213 sentences (mentalistic and mechanistic) were positioned on the two extremes of a slider. The slider was
214 associated with a 0-100 scale, which was not shown to participants (see Figure 1). The cursor was initially
215 always placed at the centre of the scale (i.e. 50). For 50% of the items, the mechanistic sentence was
216 presented on the left side of the slider, while the mentalistic was presented on the right side. For the other
217 50%, the location of mechanistic and mentalistic sentences was reversed. The order of presentation of the
218 items was randomized.
219 Data Analysis and Results
220 All statistical analyses were performed in R (version 3.4.0, freely available at http://www.rproject.org).
221 Data analysis was conducted on a sample including responses collected from 106 participants. For each
222 participant we calculated the InStance score. The score was computed as the average score of all questions.
223 Scores under 50 meant the answer was mechanistic’, scores above 50 meant they were ‘mentalistic’.
224 The overall average score for the InStance questionnaire was 40.73 (with 0 value indicating the most
225 mechanistic score, and 100 indicating the most mentalistic score). We tested the distribution of InStance
226 scores for normality with the Shapiro–Wilk test. Results showed that Average InStance scores were
227 distributed normally, W = 0.99, p > .05. We analysed scores of each item of the InStance questionnaire for
10
228 outliers, values that lie outside 1.5 * Inter Quartile Range (IQR, the difference between 75th and 25th
229 quartiles), and did not find any outlier question.
230
231 Figure 2. InStance average scores distribution (N=106).
232 The average InStance scores for each item are summarized in Table 2. Two scenarios that scored highest
233 in mentalistic descriptions (78.16 and 68.83 on average, respectively) are presented in Figure 3; two
234 scenarios with the most mechanistic scores (10.07 and 11.97 on average, respectively) are presented in
235 Figure 4.
236
11
A.
B.
237 Figure 3. Two scenarios with highest mentalistic scores. Panel A shows Item 26 which received the score 78.16 on average, while 238 panel B decipts Item 25 which received the score of 68.83 on average. Persons depicted in the photographs provided written consent 239 to be depicted in the photographs used in our questionnaire.
A.
B.
240 Figure 4. Two scenarios with highest mechanistic scores. Panel A shows Item 29 which received the score 10.07 on average, while 241 panel B depicts Item 32 which received the score of 11.97 on average.
12
242 Table 2. Average score and standard deviation for each item of the InStance questionnaire (N=106).
N° humans in N° humans in Item Mean SD Item Mean SD the scenario the scenario 1 0 55.08 44.39 19 1 38.09 38.33 2 0 35.25 39.72 20 0 24.71 31.91 3 0 46.37 41.91 21 1 56.58 39.13 4 1 50.28 41.71 22 2 24.42 33.80 5 1 31.51 35.74 23 0 34.34 37.02 6 0 31.45 38.51 24 1 52.42 42.11 7 0 43.46 40.89 25 0 68.83 38.41 8 2 33.27 38.11 26 1 78.16 30.59 9 0 22.84 30.21 27 1 57.61 38.94 10 0 55.72 41.15 28 1 45.08 40.81 11 1 32.35 37.37 29 0 10.07 20.62 12 1 66.05 39.47 30 0 48.55 41.41 13 0 28.56 35.14 31 0 25.10 34.34 14 0 42.65 40.26 32 0 11.97 25.22 15 0 41.85 40.47 33 0 34.21 38.33 16 0 33.79 38.01 34 0 46.40 41.53 17 0 33.28 36.97 35 0 34.40 36.91 18 1 65.04 38.54 243
244 In order to test internal consistency of the responses to the InStance questionnaire, we calculated Cronbach's
245 alphas, which yielded a result of .82 for 35 items, indicating high internal consistency of the InStance items.
246 To uncover possible differences in rating scores as a function of presence of a human in the scenarios (one,
247 two or zero humans) we conducted a principal-components analysis (PCA, varimax method). Results
248 showed that the three components (2 vs. 1 vs. 0 humans in the scenario) together accounted for 29% of the
249 variance (see Table 3) indicating that items involving humans did not predict different rating scores.
250
13
Table 3. Weights of the InStance scores for the three components.
Component Component Item 1 2 3 Item 1 2 3 1 -.226 .010 .277 19 .295 .144 -.003 2 .142 .066 .570 20 .191 .364 .165 3 .259 .122 .581 21 .502 -.106 -.014 4 .486 .363 .062 22 .422 .333 .044 5 -.087 .093 .353 23 .408 .153 -.078 6 .201 .307 .034 24 .560 .113 .094 7 .141 .137 .696 25 .520 -.126 .329 8 .116 .542 .138 26 .617 -.222 .072 9 -.009 .518 .032 27 .535 -.074 .057 10 .485 .288 .055 28 .500 .119 .357 11 -.075 .738 .240 29 .022 .299 .026 12 .464 .412 -.061 30 .420 -.015 -.094 13 -.072 .174 .540 31 -.108 .326 -.006 14 .401 .350 .274 32 .061 .569 .034 15 .140 .249 -.003 33 .171 .019 .524 16 -.116 .442 .376 34 .246 .317 .170 17 -.039 .305 .477 35 .062 -.212 .647 18 .558 .119 .109 251
252 We evaluated the associations between the InStance scores and gender, age, education, number of children,
253 and siblings of the respondents by a multiple linear regression. Results showed no significant associations
254 between the InStance scores and the respondent characteristics of age, gender, education, number of
255 children, and siblings (all ps > .32 ). To assess the relationship between the InStance scores and scores in
256 Reading the Mind in the Eyes and Völlm questionnaires, Pearson product-moment correlation coefficients
257 were calculated. Results showed no correlation between the InStance scores and the scores in the Reading
258 the Mind in the Eyes or Völlm questionnaires (all ps > .06 ).
259 From 106 respondents, 84% reported that they were completely unfamiliar with robots (N= 89) while the
260 remaining 16% reported various levels of familiarity (N= 17). In order to check whether the InStance scores
261 were associated with the degree of familiarity of the respondents with humanoid robots, we performed a
262 linear regression. No effect of familiarity with robots on the Instance score was observed, p=.21. The
263 average scores for the InStance questionnaire were 36.69 and 41.50 for familiar and not familiar with robots
264 respondents, respectively.
265
14
266 Analyses on the group not familiar with robots
267 In addition to the analyses of data from the entire sample, we conducted further analyses on data from
268 participants who reported no familiarity with robots. As outlined above, our main aim was to investigate
269 intentional stance for humanoids in the general population - people with no previous experience with
270 robots. For the group of respondents who were not familiar with robots, the InStance scores were also
271 distributed normally (Figure 5), Shapiro–Wilk test: W = 0.99. p > .05.
272
273 Figure 5. InStance scores distribution for the not familiar group (N=89).
274 Results showed also no significant associations between the InStance scores and the respondent
275 characteristics of age, gender, education, number of children, and siblings for the group of respondents who
276 were not familiar with robots (all ps > .16). Similarly, no correlations between the InStance scores and the
277 scores in the Reading the Mind in the Eyes and Völlm questionnaires were found (all ps >.16 ). We
278 estimated the percentage of participants who attributed ‘mechanistic’ or ‘mentalistic’ descriptions
279 according to their average InStance score. Participants were classified into two groups according to their
280 InStance score. Participants who scored below 50 (0 - 50 in our scale) were attributed to the Mechanistic
281 group (N=62), whereas participants with an Instance score above 50 (50 – 100) were classified as the
15
282 Mentalistic group (N= 27). To check whether the frequencies of the answers in each category (i.e.,
283 Mechanistic and Mentalistic) differed from the chance of level (i.e., expected frequency of 0.5), we
284 performed a chi-square test. Results revealed that both Mechanistic (69.7 %) and Mentalistic (30.3 %)
285 scores occurred more frequently than the chance level, X2(1. N = 89) = 13.76. p < .001, Figure 6. In order
286 to compare if the mean InStance scores of the two groups significantely differed from the null value of our
287 scale, we run one-sample t-tests against a critical value of 50 (i.e. the zero of our scale). Results showed
288 that the mean InStance score significantly differed from the null value of 50 both for the Mechanistic (M
289 = 34.19. SEM = 1.28. t (61) = -12.33. p < .0001) and the Mentalistic group (M = 58.27. SEM = 1.18. t (26)
290 = 7.03. p < .0001).
80% 70% 60% 50% 40% 30% 20% 10% 0% All sample Not Familiar Group Mechanistic 71,70% 69,70% Mentalistic 28,30% 30,30% 291 292 Figure 6. Percentage of Mechanistic (green bars) and Mentalistic (yellow bars) responses for the entire sample (N=106) on the left, 293 and the Not Familiar group (N=89) on the right.
294
295 General Discussion
296 The aim of the present study was to develop a method for assessing how likely it is that humans adopt the
297 intentional stance towards the humanoid robot iCub. To meet this aim, we developed a questionnaire that
298 consisted of 35 items. Each item was a sequence of three photographs depicting iCub involved in a
299 naturalistic action. Below each sequence, there was a slider scale on which participants could move the
300 cursor in the direction of one of the extremes of the scale. On one of the extremes, there was a mentalistic
16
301 description of what was depicted in the photographs, on the other extreme there was a mechanistic
302 description. We administered the questionnaire online, and received 106 responses. The high internal
303 consistency (α = .82) of mean scores of the items revealed that the questionnaire we developed was uniform.
304 This is consistent also with the result of the principal component analysis (PCA). Specifically, no difference
305 were found between “social” (iCub interacting with one or two other agents) and “individual” (iCub alone)
306 scenarios. Thus, we can conclude that our tool is reliable.
307 No correlation was found between sociodemographic characteristics and InStance ratings of participants.
308 Adopting the intentional or design stance was not affected by socioeconomics factors (i.e. education) for
309 our questionnaire. Furthermore, we found no correlation with the two tests used for the assessment of
310 Theory of Mind abilities, namely the Reading Mind in the Eyes (Golan, et al., 2006) and the Völlm comic-
311 strip questionnaire (Völlm et al., 2006). The main reason behind this might be that error rates in healthy
312 adult population in these two tests are typically very low, and, in our sample, participants were at ceiling
313 performance and there was almost no variance in terms of accuracy. Arguably, testing psychiatric
314 population could lead to a higher variance of accuracy between participants, and could provide additional
315 information regarding the relationship between Theory of Mind abilities and adoption of intentional stance
316 towards non-human agents. Importantly for our purposes, however, using these tests showed that we did
317 not have any outliers in terms of Theory of Mind abilities.
318 Overall, our results indicate that participants showed a slight bias toward a mechanistic explanation when
319 they were asked to evaluate robot actions (mean overall score = 40.73). This is not surprising given that
320 participants were observing a robot. This is also in line with previous studies emphasizing that a lower
321 degree of intentionality is attributed to machines when explaining their actions, while a higher degree is
322 attributed when perceiving conspecific’s behaviours (Krach et al., 2008; Chaminade et al., 2010; Wiese,
323 Metta, Wykowska, 2017). The results is also in line with the essence of the concept of the intentional stance
324 (Dennett, 1987). We can speculate that, in order to understand a humanoid behavior, humans are more
325 prone to adopt the design, rather than the intentional, stance – and this is in spite of the natural tendency to
326 anthropomorphize unknown entities (Epley et al., 2007). Perhaps nowadays, humanoid robots are not
327 unknown entities anymore: in the Western world, some of them are already used at airports, shops or
17
328 cultural events to provide information to visitors (Aaltonen et al., 2017; Suddrey et al., 2018). The exposure
329 Stone et al. (2017) hypothesized might entail a habituation effect. This in turn could have an impact on
330 likelihood of anthropomorphism toward humanoid robots, thus leading to a more mechanistic evaluation
331 of theirs actions. Koay and colleagues (Koay, Syrdal, Walters, Dautenhahn, 2007) demonstrated that
332 exposure to robots significantly impacts their perception by participants.
333 However, and interestingly, the design stance was not always chosen in order to explain iCub’s actions, as
334 the average score was significantly different from 0 (where 0 represents “mechanistic” explanations). This
335 clearly indicates that participants have often chosen mentalistic explanations of the given scenarios. This
336 choice could have depended on specific scenarios (some were more likely to be interpreted mentalistically
337 than mechanistically), or also on individual differences among participants (there was a sub-sample that
338 was more likely to adopt the intentional stance, and a sub-sample that was more likely to adopt the design
339 stance). This shows that in principle it might still be possible to induce adoption of intentional stance
340 towards artificial agents. The likelihood of adopting the intentional stance might depend on the context in
341 which the robot is observed, its behavioral characteristics (e.g., contingency of its behavior on participant’s
342 behavior, cf. Willemse et al., 2018), cultural background (attitude towards humanoid robots is strongly
343 associated with culture (for a rev. see Haring, et al., 2014) and also individual differences of participants.
344 In order to further investigate potential factors contributing to the choice between the mentalisitc vs.
345 mechanistic rating we asked participants whether they had previous experience with robots. Our data did
346 not show any effect of familiarity with robots (p > .05) on the ratings, but it is important to point out that
347 the majority of our sample consisted of people not familiar with this kind of technology (n = 89). To
348 understand how familiarity and expertise might affect adopting the intentional stance towards robots, we
349 conducted follow-up analyses on the sample not familiar with robots.
350 By analyzing the differences in scores between items in the not familiar group, we noticed that some
351 scenarios strongly elicited a mentalistic explanation. Interestingly, when we examined mentalistic (M =
352 58.27) and mechanistic (M = 34.19) evaluations between participants, we found that both types of scores
353 were significantly different from the middle point of our scale (for both, one-sample t-test revealed
354 significant difference from 50, p<.001). This suggests that participants were clearly choosing for each
18
355 scenario either a mechanistic or a mentalistic explanation, neither answering randomly nor relying on the
356 middle point of the scale. Thus, we can argue that the adoption of intentional or non-intentional models to
357 explain a robot behavior seems to be more of an all-or-nothing process (Keijzer, 2006). Our data suggest
358 that humans can shift from the design stance to the intentional stance (and vice versa), and that their choices
359 are polarized. In order to explore this effect further, we conducted a polynomial curve-fitting analysis on
360 the density of the raw scores. Results showed a significant quadratic trend (t= 5.59; p<.001, R2= .83),
361 supporting the polarization hypothesis (Fig. 7).
362
363 Figure 7. Plot of raw data for the not familiar group (N=89 Shapiro–Wilk test: W = 0.81 p < .001).
364 A possible explanation has been proposed by Davidson (1999). He pointed out that humans possess many
365 types of vocabulary for describing nature of mindless entities, or for describing intentional agents, but
366 might lack a way of describing what is between the two (Davidson, 1999). This is also in line with Dennett’s
367 proposal (Dennett, 1981) of intentional stance, and seems to be an adaptive mechanism: when information
368 about an entity changes, our model to explain the subsequent behavior needs to be adapted. By the same
369 token, when we find a model that is efficient to explain a behavior, we take its prototypical explanation,
370 and do not necessarily search for explanations that are in the grey fuzzy zones of in-between models. In
371 each scenario of our experiment, the robot was always the same, but the action and the environment around
372 it changed across conditions, modulating participants’ ratings. However, once there was a bias towards
19
373 either the mentalistic or the mechanistic explanation, the rating became quite polarized. This might be a
374 consequence of general tendency of people to form discrete categories rather than continuous fuzzy
375 concepts (Dietrich, & Markman, 2003). Given the flexibility in adopting mechanistic or mentalistic model,
376 we hypothesize that context variables might influence the perception of a humanoid robot’s actions in
377 intentional terms, in a similar way as it occurs for human-likeness and anthropomorphism (Fink, 2012).
378 In summary the present study used a novel method for evaluating whether intentional stance is adopted
379 towards a humanoid robot. Our results show that it is possible to induce adoption of the intentional stance
380 towards artificial agents which have some degree of human-like appearance. Understanding what stance
381 participants adopt towards an artificial agent might have consequences for social attunement with the agent,
382 and how the agent (robot) should be designed in order to optimize its acceptance. Therefore, the questions
383 we ask in our research and the methods we use will provide not only new theoretical insights for the study
384 of human social cognition, but also contribution to the design of new robot platforms that are to interact
385 with humans. Further research needs to explore what are the exact factors that influence adoption of
386 intentional stance. .In this context, it is obvious that building robots that encourage social interaction is not
387 only an engineering endeavor, but also requires a deeper understanding of social-cognitive mechanisms of
388 the human brain that occur during the human-robot interaction (Martini, 2016; Wiese, et al., 2017; ).
389 Conflict of interest
390 The authors declare that the research was conducted in the absence of any commercial or financial
391 relationships that could be construed as a potential conflict of interest.
392 Author contribution 393 SM, DG, FC, EB, AW designed the study; SM developed the experimental materials; EB and FC
394 analyzed the data; SM, DG, FC, AW wrote the paper.
20
395 Funding
396 This project has received funding from the European Research Council (ERC) under the
397 European Union’s Horizon 2020 research and innovation program (grant awarded to AW, titled
398 “InStance: Intentional Stance for Social Attunement”. ERC starting grant, grant agreement No.: 715058)
399
21
400 Acknowledgment
401 We express our gratitude to Kyveli Kompatsiari, Jairo Pérez-Osorio, Elef Schellen, Davide de Tommaso,
402 and Cesco Willemse for contributing to preparation of the InStance questionnaire items. A special thanks
403 to Kyveli Kompatsiari, Jairo Pérez-Osorio and Laura Taverna for additional help with preparing the
404 photographs. We would like to thank also AC, CB, EB, GR, HP, KK and MBA for their willingness and
405 consent to be depicted in the photographs.
406
407
22
408 References
409 Aaltonen, I., Arvola, A., Heikkilä, P., & Lammi, H. (2017, March). Hello Pepper, May I Tickle You?:
410 Children's and Adults' Responses to an Entertainment Robot at a Shopping Mall. In Proceedings of the
411 Companion of the 2017 ACM/IEEE International Conference on Human-Robot Interaction (pp. 53-54).
412 ACM.
413 Baack, S. A., Brown, T. S., & Brown, J. T. (1991). Attitudes toward computers: Views of older adults
414 compared with those of young adults. Journal of Research on Computing in Education, 23(3), 422-433.
415 Baron- Cohen, S., Jolliffe, T., Mortimore, C., & Robertson, M. (1997). Another advanced test of theory of
416 mind: Evidence from very high functioning adults with autism or Asperger syndrome. Journal of Child
417 psychology and Psychiatry, 38(7), 813-822.
418 Baron-Cohen, S., Wheelwright, S., Hill, J., Raste, Y., & Plumb, I. (2001). The “Reading the Mind in the
419 Eyes” Test revised version: a study with normal adults, and adults with Asperger syndrome or high-
420 functioning autism. The Journal of Child Psychology and Psychiatry and Allied Disciplines, 42(2), 241-
421 251.
422 Bartneck, C., & Reichenbach, J. (2005). Subtle emotional expressions of synthetic characters. International
423 journal of human-computer studies, 62(2), 179-192.
424 Breazeal, C., & Scassellati, B. (1999). How to build robots that make friends and influence people.
425 In Intelligent Robots and Systems, 1999. IROS'99. Proceedings. 1999 IEEE/RSJ International
426 Conference on (Vol. 2, pp. 858-863). IEEE.
427 Chaminade, T., Hodgins, J., & Kawato, M. (2007). Anthropomorphism influences perception of computer-
428 animated characters’ actions. Social cognitive and affective neuroscience, 2(3), 206-216.
429 Chaminade, T., Zecca, M., Blakemore, S. J., Takanishi, A., Frith, C. D., Micera, S., & Umiltà, M. A. (2010).
430 Brain response to a humanoid robot in areas implicated in the perception of human emotional gestures.
431 PLoS one, 5(7), e11577.
23
432 Chaminade, T., Rosset, D., Da Fonseca, D., Nazarian, B., Lutscher, E., Cheng, G., & Deruelle, C. (2012).
433 How do we think machines think? An fMRI study of alleged competition with an artificial
434 intelligence. Frontiers in human neuroscience, 6, 103.
435 Davidson, D. (1999). The emergence of thought. Erkenntnis, 51(1), 511-521.
436 Dennett, D. C. (1987). True believers: The intentional strategy and why it works. The Intentional Stance
437 (MIT Press, Cambridge, Mass.), pp. 13-35.
438 Dennett, D. C. (1971), Intentional Systems. The Journal of Philosophy, Vol.68, No. 4, (25 February 1971),
439 pp. 87-106.
440 Dictionary of the History of Ideas: Studies of Selected Pivotal Ideas, edited by P. P. Wiener, NY: Scribner,
441 1968, 1973, 87-91.
442 Dietrich, E., & Markman, A. B. (2003). Discrete thoughts: Why cognition must use discrete
443 representations. Mind & Language, 18(1), 95-119.
444 Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: a three-factor theory of
445 anthropomorphism. Psychological review, 114(4), 864.
446 Fink, J. (2012, October). Anthropomorphism and human likeness in the design of robots and human-robot
447 interaction. In International Conference on Social Robotics (pp. 199-208). Springer, Berlin, Heidelberg.
448 Gallagher, H. L., Jack, A. I., Roepstorff, A., & Frith, C. D. (2002). Imaging the intentional stance in a
449 competitive game. Neuroimage, 16(3), 814-821.
450 Haring, K. S., Silvera-Tawil, D., Matsumoto, Y., Velonaki, M., & Watanabe, K. (2014, October).
451 Perception of an android robot in japan and australia: A cross-cultural comparison. In International
452 Conference on Social Robotics (pp. 166-175). Springer, Cham.
453 Heider, F., & Simmel, M. (1944). An experimental study of apparent behavior. The American journal of
454 psychology, 57(2), 243-259.
24
455 Kaplan, F. (2004). Who is afraid of the humanoid? Investigating cultural differences in the acceptance of
456 robots. International journal of humanoid robotics, 1(03), 465-480.
457 Keijzer, F. (2006, January). Differentiating Animality from Agency Towards a Foundation for Cognition.
458 In Proceedings of the Annual Meeting of the Cognitive Science Society (Vol. 28, No. 28).
459 Koay, K. L., Syrdal, D. S., Walters, M. L., & Dautenhahn, K. (2007, August). Living with robots:
460 Investigating the habituation effect in participants' preferences during a longitudinal human-robot
461 interaction study. In Robot and Human interactive Communication, 2007. RO-MAN 2007. The 16th
462 IEEE International Symposium on (pp. 564-569). IEEE.
463 Krach, S., Hegel, F., Wrede, B., Sagerer, G., Binkofski, F., & Kircher, T. (2008). Can machines think?
464 Interaction and perspective taking with robots investigated via fMRI. PloS one, 3(7), e2597.
465 Golan, O., Baron-Cohen, S., Hill, J.J. and Golan, Y. (2006) The 'Reading the Mind in Films' Task:
466 complex emotion recognition in adults with and without autism spectrum conditions Social
467 Neuroscience, 1(2): 111-123. Doi: 10.1080/17470910600980986.
468 Gonzàles, A., Ramirez, M. P., & Viadel, V. (2012). Attitudes of the elderly toward information and
469 communication technologies. Educ. Gerontol, 38(9), 585-594.
470 Gould S. J. (1996). Can we truly know sloth and rapacity? Nat. Hist. 105 18–57.
471 Martini, M. C., Gonzalez, C. A., & Wiese, E. (2016). Seeing minds in others–Can agents with robotic
472 appearance have human-like preferences? PloS one, 11(1), e0146310.
473 Metta, G., Sandini, G., Vernon, D., Natale, L., & Nori, F. (2008). The iCub humanoid robot: an open
474 platform for research in embodied cognition. In Proceedings of the 8th workshop on performance
475 metrics for intelligent systems (pp. 50-56). ACM.Morewedge, C. K., Preston, J., Wegner, D. M.,
476 Timescale Bias in the Attribution of Mind. Journal of Personality and Social Psychology, Vol 93(1), Jul
477 2007, 1-11
478 Nagel, T. (1974). What is Like to be a Bat. The Philosophical Review. 83 (4): 435-450. Doi:
479 10.2307/21183914.
25
480 Parlangeli, O., Chiantini, T., & Guidi, S. (2012). A mind in a disk: The attribution of mental states to
481 technological systems. Work, 41(Supplement 1), 1118-1123.
482 Scholl, B. J., & Tremoulet, P. D. (2000). Perceptual causality and animacy. Trends in cognitive
483 sciences, 4(8), 299-309.
484 Serafin, M., & Surian, L. (2004). Il Test degli Occhi: uno strumento per valutare la" teoria della
485 mente". Giornale italiano di psicologia, 31(4), 839-862.
486 Stone, P., Brooks, R., Brynjolfsson, E., Calo, R., Etzioni, O., Hager, G., ... & Leyton-Brown, K. (2016).
487 Artificial intelligence and life in 2030. one hundred year study on artificial intelligence: Report of the
488 2015-2016 Study Panel. Stanford, CA: Stanford University. http://ai100. stanford. edu/2016-report.
489 Accessed February, 7, 2017.
490 Suddrey, G., Jacobson, A., & Ward, B. (2018). Enabling a Pepper Robot to provide Automated and
491 Interactive Tours of a Robotics Laboratory. arXiv preprint arXiv:1804.03288.
492 Völlm, B. A., Taylor, A. N., Richardson, P., Corcoran, R., Stirling, J., McKie, S., & Elliott, R. (2006).
493 Neuronal correlates of theory of mind and empathy: a functional magnetic resonance imaging study in
494 a nonverbal task. Neuroimage, 29(1), 90-98
495 Waytz, A., Epley., N., Cacioppo, J. T. (2010). Social Cognition Unbound: Insights Into Antropomorphism
496 and Dehumanization. Curr. Dir. Psychol Sci., 9 (1): 58-62. Doi: 10.1177/0963721409359302.
497 Wiese, E., Wykowska, A., Zwickel, J., & Müller, H. J. (2012). I see what you mean: how attentional
498 selection is shaped by ascribing intentions to others. PloS one, 7(9), e45391.
499 Wiese, E., Metta, G., & Wykowska, A. (2017). Robots as intentional agents: Using neuroscientific methods
500 to make robots appear more social. Frontiers in psychology, 8, 1663.
501 Willemse, C., Marchesi, S., & Wykowska, A. (2018). Robot Faces that Follow Gaze Facilitate Attentional
502 Engagement and Increase Their Likeability. Frontiers in Psychology, 9, 70.
503 Wykowska, A., Kajopoulos, J., Ramirez-Amaro, K., Cheng, G. (2015a). Autistic traits inversely correlate
504 with implicit sensitivity to human-like behavior. Interaction Studies, 16:2, 219-248.
26
505 Wykowska, A., Kajopoulos, J., Obando-Leitón, M., Chauhan, S., Cabibihan, J.J., Cheng, G. (2015b).
506 Humans are well tuned to detecting agents among non-agents: Examining the sensitivity of human
507 perception to behavioral characteristics of intentional agents. International Journal of Social Robotics,
508 7 (5), 767–781, doi: 10.1007/s12369-015-0299-6
509 Wykowska, A., Wiese, E., Prosser, A., & Müller, H. J. (2014). Beliefs about the minds of others influence
510 how we process sensory information. PLoS One, 9(4), e94339.
511 Zickuhr, K., & Smith, A. (2012). Digital differences. PewResearch Center, Report.
512
27