Artificial Intelligence Promotes Children’s Science Learning from Television Shows Through Enhanced Verbal Engagement

• In Event: AI in Early Childhood: Impacts on STEM and Language Learning

Abstract

Objective
Video programs are important and accessible educational resources, and their potential can be amplified if children are allowed to interact with media characters during their video watching. Partnering with PBS KIDS, we developed interactive videos based on a popular show, Elinor Wonders Why. In these videos, the main character, Elinor, powered by a conversational agent, engaged children in conversation by asking questions and providing responsive feedback (Figure 1). To examine the impact of these interactive videos on children’s learning, we conducted an experiment involving 240 children, demonstrating the effectiveness of using conversational technologies to enhance the educational benefits of television watching. Additionally, we investigated children’s verbal engagement with interactive videos and its mediating effect on children’s learning.

Theoretical Framework
Television programs have traditionally been a one-way transmission, lacking opportunities for children to actively interact with the educational content (Jing & Kirkorian, 2020a). This limitation can be mitigated through dialogic interaction, where children are asked questions and receive responsive feedback, enhancing learning from video content (Anderson & Hanson, 2017). A crucial mechanism in dialogic interactivity that enhances learning is children’s verbal engagement (Xu et al., 2021), captured by the quantity (e.g., response rate) and quality (e.g., relevance) of their responses.

Method
We randomly assigned 240 children into one of the three conditions: 1) watching the interactive episode, 2) watching the pseudo-interactive episode (in which the media character asks children the same questions and gives generic feedback after a fixed amount of time), or 3) watching the broadcast version of the episode. Each child watched a total of two episodes in their assigned condition, after each of which they received immediate posttests.

Results
Results are summarized in Figure 2, 3, & 4.

Children who watched the interactive video scored 0.32 SD higher than those who watched the broadcast video (p < 0.001), and 0.16 SD higher than those who watched the pseudo-interactive version (p < 0.05).

Children in the interactive condition had a 0.38 SD higher response rate than those in the pseudo-interactive condition on initial attempts to answer questions (p < .05). This difference increased to 0.50 SD with additional scaffolding and corrective feedback for incorrect responses given to children in the interactive condition (p < .01).

In the interactive condition, children’s odds of responding increased by 5.0% with each question (OR = 1.050, p < .001), whereas in the pseudo-interactive condition, the odds decreased by 0.5% per question (OR = 0.995, p < .001), with a significant interaction effect (β = -0.054, p < .001).

Children in the interactive condition exhibited increased response rate to more challenging questions (β = 0.89, p < .001), which contributed to science learning (β = 0.11, p < .05), compared with children in the pseudo-interactive condition.

Scholarly Significance
As children spend more time watching videos and increasingly use Internet-connected devices, it is crucial to explore how new forms of videos can enhance learning. Researchers and designers should leverage advancing conversational technologies, such as generative AI, to optimize the educational value of children’s screen time.

Authors

• Ying Xu, Harvard University

• Kunlei He, University of California - Irvine

• Julian Levine, University of California - Irvine

• Daniel Ritchie, University of California - Irvine

• Echo Zexuan Pan, Harvard University

• Andres Sebastian Bustamante, University of California - Irvine

• Mark Warschauer, University of California - Irvine