Younger, not older, children have a bias to trust human more than robot informants

• In Event: Children’s selective trust across domains, agents, and cultures

Abstract

One important source of early learning in children is through observation or interaction with other individuals. Previous research suggests that humans, compared to non-humans, may be a privileged source of information for early social learning. For example, 12-month-old infants showed enhanced learning of an object that was looked at by a human but not by a robot (Okumura et al., 2013). Three-year-olds appeared to have a robust bias to trust what people say (Jaswal et al., 2010). However, such a predisposition to treat human testimony veridically, possibly over others, may change with age as children accumulate experience of (unreliable) testimony. By 4 years of age, for example, children not only showed a preference to learn from accurate over incorrect human informants (Tong et al., 2019), but they also applied similar selective trust toward robot informants (Brink & Wellman, 2020). Yet, little is known about the developmental change in children’s selective trust across human and robot informants. Here, we explore the developmental differences between 3- to 5-year-old children in their willingness to learn novel labels from a human or robot informant with a history of accuracy or inaccuracy.
Participants were 120 preschoolers (M=4.56 years, SD=0.59, range=3.50-5.54; 57 girls). Children were randomly assigned to one of four conditions (n=30 per condition), where the informant-type (either a human adult or a NAO robot) and the accuracy of the informant (either accurate or inaccurate) were manipulated across conditions. Children watched video clips of a single informant labeling familiar and unfamiliar objects on a computer. During a history phase, children viewed the same informant labeling a series of familiar objects (i.e., apple, ball, book) either consistently accurate or consistently inaccurate. Following that, children went through four test trials concerning four novel objects, where the informant provided a novel label for each novel object. We examined whether children would accept the novel object label in the test trials.
We conducted binomial mixed-effect models to estimate children’s responses in favor of the informant’s testimony with three categorical predictors: age (younger vs. older; median-split based on Md=4.58), informant-type, and accuracy. Results revealed significant 2-way interactions of Age*Accuracy (beta=-14.20, p<.001) and Informant-type*Accuracy (beta=-13.64, p=.004). Importantly, we found a significant 3-way interaction (beta=13.58, p=.019). Further analyses found a significant Informant-type*Accuracy interaction only in the younger, not older, group (Figure 1).
Younger children were equally accepting new information from accurate human and robot informants but discriminating in accepting new information from inaccurate human and robot informants, i.e., children in the inaccurate human condition tended to accept the novel labels above chance, but not those in the inaccurate robot condition. In contrast, older children reliably trusted the accurate informant and mistrusted the inaccurate informant, regardless of whether the informant was a human or a robot.
Our results suggested a developmental change in preschoolers’ reliance on informants’ epistemic characteristics to inform their selective trust decisions, especially human informants. This developmental shift could be a result of a reduction of children’s bias to treat humans as privileged sources of information as they get older.

Authors

• Xiaoqian Li, Singapore University of Technology and Design
  Presenting Author

• W. Quin Yow, Singapore University of Technology and Design