Abstract:

A new generation of interactive models, called digital voice assistants (DVAs), can respond to young children's speech requests automatically and interact with them by voice. Research on the development of young children's epistemic trust in DVAs is scarce. Previous research has concentrated on the development and influencing factors of young children's epistemic trust in human informants or traditional electronic media (e.g., computers, webpages, internet). The semisocial nature of these devices determines the specific theoretical and practical value of investigating young children's epistemic trust in DVAs. Based on this, the purpose of the current study was to investigate the epistemic trust of young children (aged 3-6) and adults in DVAs in various domains and to confirm the significance of accuracy in their trust.

The paradigm of dual-informant sources was employed in both experiments. A sample size of 88 children was required for an effect size of w = 0.30, 1− β = 0.8, α = 0.05, according to G*Power 3.1. In Experiment 1, 30 adults and 90 children aged 4-6 were given testimony from distinct information sources (DVAs vs. humans) in either the natural or social domain to investigate the children's willingness to ask questions, explicit trust judgments, and final endorsements. Whereas the natural domain involved a task to label novel things, the social domain involved inquiry into social customs. The accuracy of the informants was manipulated in Experiment 2, which was based on Experiment 1, and 90 children aged 3-5 and 30 adults were exposed to various informants.

The research participants were asked questions about their willingness to ask, explicit trust judgments, and final endorsements. Children's and adults’ mean (SD) willingness to ask, explicit trust judgments, and final endorsements scores in each age group and domain in Experiment 1 are presented in Table 1.

A GLMM (generalized linear mixed model) model was developed with the willingness to ask as the dependent variable, and the results showed there was a significant main effect of the domain (χ² = 3.85, df= 1, p = 0.050). Post hoc comparisons showed that children were significantly less likely to choose to ask about DVAs in the social domain than in the natural domain (t= 2.19, p= 0.029 < 0.05, d = 0.14). There was also a main effect of age (χ² = 60.12, df= 3, p < 0.001). Paired comparisons showed that adults were more likely to ask about DVAs than young children (adults and 4-year-olds: t= 9.39, p< 0.001, d = 0.93; adults and 5-year-olds: t= 7.10, p< 0.001, d = 0.72; adults and 6-year-olds: t= 5.11, p< 0.001, d = 0.54) and children were more likely to ask about DVAs as they get older (5-year-olds and 4-year-olds: t= 2.30, p= 0.022 < 0.05, d = 0.19; 6-year-olds and 4-year-olds: t= 4.28, p< 0.001, d = 0.36; 6-year-olds and 5-year-olds: t= 1.98, p= 0.048 < 0.05, d = 0.17).

A GLMM model was developed with explicit trust judgments as the dependent variable, and the results showed there was a significant main effect of age (χ² = 32.37, df= 3, p < 0.001). Post hoc comparisons showed that adults were more likely to trust DVAs than young children (adults and 4-year-olds: t= 10.48, p< 0.001, d = 1.12; adults and 5-year-olds: t= 10.59, p< 0.001, d = 1.13; adults and 6-year-olds: t= 6.62, p< 0.001, d = 0.79) and 6-year-olds were more likely to trust DVAs than 4- and 5-year-olds (6-year-olds and 4-year-olds: t= 3.86, p< 0.001, d = 0.31; 6-year-olds and 5-year-olds: t= 3.97, p< 0.001, d = 0.32).

A GLMM model was developed with final endorsements as the dependent variable, and the results showed there was a significant main effect of age (χ² = 33.49, df= 3, p < 0.001). Post hoc comparisons showed that adults were more likely to finally trust DVAs than young children (adults and 4-year-olds: t= 10.67, p< 0.001, d = 1.15; adults and 5-year-olds: t= 10.46, p< 0.001, d = 1.13; adults and 6-year-olds: t= 7.41, p< 0.001, d = 0.86) and 6-year-olds were more likely to finally trust DVAs than 4- and 5-year-olds (6-year-olds and 4-year-olds: t= 3.27, p= 0.001, d = 0.26; 6-year-olds and 5-year-olds: t= 3.05, p= 0.002 < 0.01, d = 0.25).

The results of Experiment 1 showed that the children preferred to ask the DVAs questions about the natural domain rather than the social domain, with the DVAs being preferred overall. Moreover, the 6-year-old children preferred the DVAs as the information source more than the 4- to 5-year-old children. The adults were more likely to trust the DVAs than the young children.

Children's and adults’ mean (SD) willingness to ask, explicit trust judgments, and final endorsements scores in each age group, accuracy condition, and domain in Experiment 2 are presented in Table 2.

A GLMM model was developed with the willingness to ask as the dependent variable, and the results showed there was a significant main effect of the accuracy condition (χ² = 97.57, df= 1, p < 0.001). Post hoc comparisons showed that children were more likely to ask for accurate informants (t= 28.62, p< 0.001, d = 1.85). There was also a significant interaction of accuracy condition and age (χ² = 25.66, df= 3, p < 0.001) (see Figure 1).

A GLMM model was developed with explicit trust judgments as the dependent variable, and the results showed there was a significant main effect of the accuracy condition (χ² = 103.29, df= 1, p < 0.001). Post hoc comparisons showed that children were more likely to trust accurate informants (t= 41.27, p< 0.001, d = 2.67).

A GLMM model was developed with final endorsements as the dependent variable, and the results showed there was a significant main effect of the accuracy condition (χ² = 88.39, df= 1, p < 0.001). Post hoc comparisons showed that children were more likely to endorse accurate informants finally (t= 42.79, p< 0.001, d = 2.76).

The results of Experiment 2 revealed that the children of all ages and adults were more likely to accept correct informant testimony in both the natural and social domains. In other words, the children were more likely to use the current accuracy of informants as a cue to assess and decide which informant to trust, and when the DVAs lost their accuracy, the children's preference disappeared along with their intellectual trust. The preference for accurate informants was more obvious in the adults and 4- to 5-year-olds than in the 3-year-olds, with the 3-year-olds being less sensitive to accuracy. Accuracy was an essential indicator of the DVAs' dependability.

Our study is the first to investigate the development of epistemic trust in DVAs among children aged 3-6 in China. The results show that children can use DVAs as a source of information and knowledge. Young children become more likely to believe the testimonies of DVAs as they grow older. Children are more likely to trust DVAs in the natural domain than in the social domain. Furthermore, young children are more likely to accept the testimony of reliable informants. The results of this study may contribute to our understanding of the usability and utility of human interaction with technological systems and offer suggestions for the use of DVAs in homes and classrooms to support early learning.

Key words: 3- to 6-year-olds, epistemic trust, digital voice assistants (DVAs), accuracy, domain specificity