Edited by: Ilaria Grazzani, University of Milano-Bicocca, Italy
Reviewed by: Flavia Lecciso, University of Salento, Italy; Antonella D’Amico, University of Palermo, Italy
This article was submitted to Developmental Psychology, a section of the journal Frontiers in Psychology
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Studying trust in the context of human–robot interaction is of great importance given the increasing relevance and presence of robotic agents in the social sphere, including educational and clinical. We investigated the acquisition, loss, and restoration of trust when preschool and school-age children played with either a human or a humanoid robot
One of the challenges of contemporary robotics is
In psychology, trust can be described as “a multidimensional psychological attitude involving beliefs and expectations about the reliability of the trustee resulting from social experiences involving uncertainty and risk” (
The establishment of trusting relationships is critical for effective interpersonal dynamics. This is particularly relevant where children are called to build new relationships with peers, educators, and other adults. An example of the importance of the construction of interpersonal trust is highlighted in a study with children under protection services (
Human trusting relationships are also shaped by past relational histories, originating with primary caregivers (e.g.
Likewise, the development of the individual’s cognitive competencies is important, particularly for the definition of the informant’s epistemic reliability. Cognitive skills allow individuals to reason about the other’s perspective and to objectively evaluate informational access. In this respect, the development of a ToM enabling individuals to conceptualize the mental states that guide behavior (
In relation to human–robot interaction, studies that have specifically investigated trust in a robot agent or system have typically involved adult participants. These studies have either used explicit measures of trust assessment, mostly involving self-reports (e.g.
In the present study, trust was explored through a novel Trusting Game (TG) named “Guess where it is” requiring the interactive partner (either the human or the robot) and, subsequently, the child to guess the position of a doll hidden under a box. Through the structure of the game, we set the conditions for the child to consequentially make the same decisions as the play partner, thus ultimately establishing a trusting relationship (e.g.
To make the child perceive the robotic agent NAO as a real interactional partner, it was introduced to children in a preliminary session when they were familiarized with some of the robot’s physical and social skills (walking, moving its arms, talking, greeting, etc.) (see
The children’s perception of the robot’s mental qualities as compared to the human was evaluated through the Attribution of Mental States (AMS) questionnaire (
Ninety-four (94) Italian kindergarten and school-age children participated in the experiment. The children were divided into four age groups as follows: 3-year-olds (
The children were assessed in two experimental sessions on different days within a 2-week time frame. In the first session, the children were administered the following tests: AMS scale (inspired by the work of
The AMS scale is a measure of the mental states that participants attribute when looking at pictures depicting specific characters, in this case a human and the robot NAO. The scale is an
The TG was inspired by the work of
Overview of the experimental setup in the Trust Game.
The TG involves three independent phases. The first phase [Trusting Acquisition (TA)] aims to assess the participant’s acquisition of trust in the other player by calculating how many trials elapse before the child follows the other player’s guess. Trust is assumed when the child follows the other player’s guess on three consecutive trials. After trust acquisition, the game switches into the second phase [Mistrust Acquisition (MA)], which assesses the participant’s acquisition of mistrust in the other player by calculating how many trials it takes for the child not to follow the other player’s guess. Mistrust is assumed when the child does not follow the other player’s guess on three consecutive trials. The last phase [Trusting Restoration (TR)] shares the same play structure as the initial phase. The game lasted, on average, between 10 and 20 min.
Each phase consisted of a maximum of 10 trials and ended after trust acquisition (phase 1), mistrust acquisition (phase 2), and trust restoration (phase 3). The switch to the following phase also occurred if the participant completed 10 trials within a given phase without completing the three-trial sequence. The dependent variable (DV) was the number of trials the child required before acquiring trust or mistrust. For example, in the initial phase, if the child started to follow the other player for three consecutive trials after the second trial (i.e. 0 0 1 1 1), the participant scored 2. If the child displayed trust immediately (i.e. 1 1 1), s/he scored 0. If the child completed the 10 trials within each phase without acquiring trust or mistrust, she/he scored 8, which is the maximum value that could possibly be attributed before ending the phase with a three-trial sequence. To compare data in the analyses, trust and trust restoration indexes were reversed to indicate, alongside trust loss, a comparable measure of the tendency to trust. Thus, a child could score between 0 (low trust) and 8 (high trust).
For the treatment of missing cases, we considered mean, median, and mode values, as well as children’s most common response patterns. The median was ultimately chosen as the most representative index for replacing missing values. Accordingly, two children were recovered for age groups 3, 5, and 9 years; one child was recovered for the age group 7 years. When an entire session was missing, the values were
The Unexpected Transfer task (
The development of a second-order false belief competence was assessed through the Ice-Cream Van task (
The Separation Anxiety Test is a semi-projective task that evaluates the child’s mental representation of his/her attachment to the caregiver. The original version developed by
The coding reflects three dimensions: (1) attachment, i.e. the ability to express vulnerability and need; (2) self-confidence, i.e. the ability to autonomously face separation; and (3) avoidance, i.e. the propensity to speak about the separation. Participants score 1 for each dimension. The final score is the result of the sum of the scores in the attachment scale and in the self-confidence scale, and of the sum of the inverse of the avoidance scale, calculated by subtracting this score from the total amount potentially obtainable on this scale. Scores range from 6 to 36, with higher scores reflecting greater quality of attachment relationship.
Children aged 3 and 5 years were administered the DCCS assessing the capacity to switch responses [for a full description of the test, please refer to
On a day that preceded the main experimental session, children were introduced to three play partners (two humans – a boy and a girl – and the robot) through video clips displayed in class on a large projector. In the videos, each of the potential partners said the same sentence: “Hello, my name is. I will be playing with you in the next days. See you soon. Bye.” The videos represented the actors while exiting a room and waving their hand to say goodbye. In this way, the children saw that the robot NAO could walk, talk, and move its head and arms.
The children were tested individually in a quiet room in their kindergarten or school. Tests were carried out by two researchers both in the morning and in the afternoon during normal activity. In the first session, the administration of the battery lasted approximately 20–30 min, depending on the child’s age. The administration of the task in the second session took about 35–45 min.
The first session started with the administration of AMS. The five AMS state categories (Perceptive, Emotional, Desires and Intentions, Imaginative, and Epistemic) were randomized across children. Afterward, children participated in the TG. At this point, the partner (i.e. human or robot) entered the experimental room and was introduced by the experimenter by his/her name: “Do you remember, this is …”. Then, both the child and the partner were invited to sit on the ground on a plastic carpet in front of an
After the game, the child was administered one of the two first-order ToM tasks and, starting from 5 years of age upward, one of the two second-order ToM tasks. The order of the ToM tasks was randomized across children, so that those who performed, for example the unexpected transfer task in the first session, completed the unexpected content task in the second session. The same was true for the second-order ToM tasks. Finally, children were given two further assessments: SAT and executive function.
Statistical analysis was carried out using the IBM Statistical Software Platform SPSS (v. 19.0). To evaluate possible differences in children’s tendency to trust the human and robot partner as a function of the child’s age and trust phase (acquisition, loss, and restoration), a repeated measures General Linear Model (GLM) analysis was carried out. The DV was the number of trials until children followed their partner (trust acquisition), stopped following their partner (trust loss), and again followed their partner (trust restoration) during the TG. To compare data from the three phases, trust and trust restoration indexes were reversed to indicate, together with trust loss, a comparable measure of the
Additionally, correlation analyses (Pearson’s
Finally, to assess possible differences in children’s mental states attribution to the robot with respect to the human partner, a repeated measures GLM analysis comparing AMS scores between human and robot was carried out as a function of the children’s age. For all the GLM analyses, the Greenhouse–Geisser correction was used for violations of Mauchly’s Test of Sphericity,
The GLM analysis, with three levels of
The results revealed a main effect of
Trust scores for the Trusting Game. Children’s average tendency to trust during the Trusting Game
Having found a consistent correlation across ages between first-order ToM and performance in the TG as described below, a further GLM was carried out using first-order ToM as a covariate. This analysis revealed a main effect of
As shown in
Association between Trust and SAT.
3 years (17) | 0.429 | –0.236 | 0.422 | 0.161 | –0.053 | 0.14 | –0.132 | ||
Acquisition | 5 years (20) | –0.042 | –0.227 | 0.088 | –0.205 | 0.282 | 0.007 | –0.004 | 0.1 |
7 years (22) | 0.2 | 0.008 | –0.264 | 0.302 | –0.135 | 0.063 | –0.12 | –0.06 | |
9 years (23) | –0.118 | –0.306 | 0.259 | 0.412 | –0.003 | 0.056 | 0.224 | ||
Overall | –0.059 | 0.027 | 0.156 | –0.172 | 0.097 | –0.025 | 0.102 | –0.056 | |
3 years (17) | –0.248 | 0.393 | 0.353 | –0.267 | 0.272 | ||||
Loss | 5 years (20) | –0.195 | –0.012 | 0.023 | –0.095 | 0.198 | –0.025 | 0.138 | –0.014 |
7 years (22) | 0.066 | 0.068 | 0.412 | 0.188 | 0.153 | 0.247 | –0.021 | ||
9 years (23) | –0.177 | –0.273 | –0.262 | 0.048 | 0.209 | –0.106 | 0.187 | ||
Overall | –0.146 | 0.092 | –0.14 | 0.07 | 0.158 | 0.052 | 0.007 | ||
3 years (17) | 0.298 | 0.418 | 0.039 | 0.163 | 0.459 | –0.307 | 0.48 | ||
Restoration | 5 years (20) | 0.129 | 0.093 | –0.147 | 0.264 | 0.362 | 0.079 | –0.118 | 0.33 |
7 years (22) | –0.17 | 0.044 | 0.05 | –0.106 | 0.269 | 0.22 | –0.016 | 0.139 | |
9 years (23) | –0.146 | –0.073 | 0.068 | –0.188 | 0.077 | 0.009 | 0.109 | 0.084 | |
Overall | –0.066 | 0.128 | 0.05 | 0.005 | 0.203 | 0.205 | –0.08 | 0.209 |
For 9-year-olds, the results also showed a positive relationship between trust acquisition and the SAT sub-dimension of attachment, indicating that more securely attached children tended to acquire trust quicker. Additionally, among 9-year-olds, there was a positive correlation between the tendency to trust during the trust loss phase and the SAT sub-dimension of avoidance. This correlation was also significant across ages.
A positive correlation was finally found between the SAT sub-dimension of attachment and the tendency to trust in the
The scores on the two ToM tasks were merged into one single score for each level of complexity (first and second order). A low level of ToM performance (coded 0) included children who scored 0 (failed) on both tasks, whereas a high level of performance (coded 1) included children who passed at least one ToM task at each complexity level.
ToM descriptives.
3 (22) | 68 | 32 | – | – |
5 (24) | 25 | 75 | 50 | 50 |
7 (24)* | 0 | 96 | 20 | 76 |
9 (23) | 0 | 100 | 13 | 87 |
All correlations found between the tendency to trust and ToM scores were negative. Thus, greater ToM abilities were associated with a lower tendency to trust, i.e. with a more reflective tendency to trust. This relationship was independent of the partner’s agency (human or robot) or the child’s age. The tendency to trust was often significantly correlated with first-order ToM, which was therefore included as a covariate in the GLM model described above. Finally, a substantial negative correlation between the tendency to trust and second-order ToM was observed during the acquisition of trust for children aged 7 years when playing with the human. These statistics are reported in
Association between Trust and ToM.
1 – Acquisition | 3 years (21) | –0.225 | − | –0.306 | − |
5 years (24) | 0.091 | 0.071 | –0.231 | –0.187 | |
7 years (23) | 0.029 | ||||
9 years (23) | –0.176 | –0.03 | |||
Overall (90/69) | –0.235 | –0.144 | |||
2 – Loss | 3 years (21) | –0.28 | − | − | |
5 years (24) | 0.079 | –0.356 | –0.303 | –0.285 | |
7 years (23) | –0.267 | 0.004 | |||
9 years (23) | –0.091 | 0.112 | |||
Overall (90/69) | –0.09 | ||||
3 – Restoration | 3 years (21) | –0.119 | − | –0.329 | − |
5 years (24) | 0.024 | –0.151 | –0.033 | 0.058 | |
7 years (23) | 0.358 | 0.317 | |||
9 years (23) | –0.298 | –0.019 | |||
Overall (91/70) | –0.163 | –0.066 | 0.143 |
Children aged 3 and 5 years were administered the DCCS, which assesses the capacity to switch between responses (
Significant age-related positive relationships were found between the ability to switch and the tendency to trust during the restoration phase among 3-year-olds when playing with the human,
A repeated measures GLM analysis comparing AMS scores between human and robot, with five levels of
Children’s scores on the Attribution of Mental States (AMS) scale. AMS mean scores for the human (HB = blue bar) and the robot (RB = orange bar) for each age group (3-, 5-, 7-, and 9-year-olds) as a function of state (Perceptual, Emotions, Intentions and Desires, Imagination, Epistemic). The bars represent the standard error of the mean. * indicates significant differences.
Exploring the three-way interaction, the most consistent difference was for the attribution of perception (HB > RB), which was significant for all four age groups,
The present study investigated trust dynamics when children aged 3, 5, 7, and 9 years played a TG
To better understand age changes in trust, the results for quality of attachment relationships, false belief understanding, and executive function skills were examined. It has been previously shown that children aged 3 and 4 years are likely to endorse information provided by someone who proved accurate in the past (see also
The development of a fundamental cognitive ability makes a substantial contribution to trust dynamics in child–robot interaction across all age groups. According to our findings, the development of ToM appears to temper the relation between quality of attachment relationships and trust, by introducing into the trust matrix a mentalistic evaluation of the other’s judgment based on an awareness of her/his/its beliefs. More specifically, children who had developed at least first-order ToM also knew that the other player did not know the position of the doll, and was therefore an unreliable informant. Not by chance, the effect of ToM on trusting behavior was most evident at 3 and 7 years of age, typically marked by the development of increasingly complex levels of ToM. When children start developing the concept of the other’s mind, they are also able to evaluate whether the other (either a human or a robot) is trustworthy on the basis of informational access. Preferential trust in either agent then moderates.
The dichotomy found between the younger and older age groups in the AMS to the robot and the human further helps to delineate the children’s perception of the robot as a mentalistic agent: For younger children, the robot is perceived as more mentalistically comparable to the human than for older children. Nevertheless, when younger children decided to trust a play partner, the affective component prevailed over the more “cool” mentalistic component, defining the preferred relational target accordingly (i.e. the human). On the other hand, the trust attributed to the robot by older children may stem from the dominance of a cognitive over an affective engagement.
In support of agent-specific differences in trusting behavior between 3- and 7-year-olds, the results also revealed a positive association between the ability to switch and trusting behavior during the trust restoration phase among children aged 3 and 7 years. Strikingly, and consistent with the data discussed above, these relations were specific to playing with the human partner for the 3-year-olds, and to playing with the robot for the 7-year-olds. In general, these correlations indicate that a greater tendency to recover trust in the other is associated with the development of the ability to switch. The specificity related to the play partner’s agency further underlines the relevance of the interactive partner for the child and reflects children’s engagement with one or the other player: 3-year-olds’ selective trust in the human was plausibly influenced by the quality of attachment relationships – as also evidenced by data on attachment described above; 7-year-olds preferential trust in the robot was possibly due to an emerging familiarity with artificial devices typical of this age. These results shed light on previous findings (
The present study provided some insight into the dynamics of trust both when relating to a human and a robot partner. Our results highlighted the impact of cognitive development, as well as children’s attachment history. We found that cognition and attachment operated separately (given the absence of a direct correlation between these two dimensions) on the establishment of trust. Particularly for children aged 3 years, trust appears to be significantly influenced by the affective dimension of trust, especially when interacting with a human. Interestingly, although securely attached children exhibited a greater tendency to trust the human, they also shifted their trust more rapidly in the trust restoration phase with the robot. This may be due to the lack of any affective bond with the robot and to the child’s cool relational attitude toward it. Effectively, this would render the robot a more “forgivable” partner.
Also, the development of false belief understanding proved to play a significant role in the establishment of trusting relationships. In particular, the development of mentalizing abilities enabled children to reflect rationally on the fact that the other player had exactly the same guessing opportunities as they did, and was therefore as susceptible to making mistakes as they were. This moderated the effect of the affective component of trust.
In the present study, the robot proved to be less susceptible to the dynamics associated with the quality of attachment relationships, and thus became a more stable trusted partner. For this reason, and particularly for children with fragile affective relational histories who have difficulties with trust, the robot might fulfill a significant scaffolding role in human–human interaction. However, an evolution of the robot as a social partner is also to be expected. Therefore, different relational dynamics may be anticipated, according to which, perhaps, an affective relation history will be created with this new entity. In this respect, a longitudinal study would further delineate the development of trust in the robot increasing the robustness of the findings. Also, a larger sample size would eventually confirm the observed tendencies.
Last, but not least, the findings from this study may inform disciplines such as Developmental Robotics on how cognitive architectures can be modeled in the robot so as to make it trusting in the human partner in a “human-like” fashion, as discussed above. This circular behavior would make the human–robot relationship increasingly ecological and, ultimately, trustful. Starting, for example, from the architectural model designed by
All data needed to evaluate the conclusions in the article are present in the article.
The studies involving human participants were reviewed and approved by the Ethic Committee, Università Cattolica del Sacro Cuore, Milano, Italy. Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.
AM, AC, PH, DM, CD, and FM conceived and designed the experiment. FM and GP conducted the experiments in schools. AM and FM secured ethical approval. CD and DM carried out the statistical analyses. AC granted the use of the SoftBank Robotics NAO humanoid robot. All authors contributed to the writing of the manuscript.
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.
We wish to thank Dr. Gabriella Fumagalli, Principal of the “I.C. Leonardo da Vinci”, and Mrs. Fulvia Maria Frigerio for their deep involvement in the organization of and support to the study. We wish to thank Dr. Mario Uboldi, Principal of the “I.C. Giovanni Pascoli”, and Mrs. Maura Panzone for their important contribution to the organization of the project and support to the study. We also wish to thank all the teachers, who actively helped in conducting the study and, in particular, the schools “Scuola dell’Infanzia Premenugo,” “Scuola dell’Infanzia Rodano,” “Scuola dell’Infanzia Caleppio,” “Scuola Primaria Rasori,” and “Scuola Primaria Settala.”