Learning with Generative Artificial Intelligence Through Co-Creative Storytelling

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

Abstract

Objective
Generative artificial intelligence (AI) holds significant potential as a learning partner in early childhood education. To leverage the capabilities of generative AI in producing free-form conversations, we developed a GPT-4-based system that can co-create stories with young children while incorporating mathematical vocabulary into the narrative. This study aims to investigate whether generative AI can support children’s learning of mathematical vocabulary through co-creative storytelling.

Theoretical Framework
Storytelling has emerged as an effective intervention to support children’s learning (Casey et al., 2004; O’Byrne et al., 2018; Rahiem, 2021). Generative AI, with its adaptive conversational capabilities, can enhance the educational benefits of storytelling by serving as an interactive scaffold (Fan et al., 2024; Ye et al., 2024). This design aligns with Vygotsky’s Zone of Proximal Development (ZPD), which underscores the role of social interaction and scaffolding in learning (Vygotsky, 1978). Additionally, theories of language acquisition highlight the importance of meaningful dialogue in the learning process (Mercer & Howe, 2012).

Method
We conducted a randomized controlled trial with 119 children (55.46% female) aged four to nine years. The children were assigned to one of three conditions where they co-created stories with different partners: a generative AI speaker (n = 41), a human experimenter face-to-face (n = 39) (i.e., present human), or a human experimenter concealed from their view (n = 39) (i.e., hidden human). An overview of the three conditions is displayed in Figure 1. During the co-creative storytelling sessions, six mathematical terms (sum, estimate, add, subtract, equal, half) were explicitly taught within narratives. Children’s learning of mathematical vocabulary was assessed through four dimensions: definition, recall, transfer, and practice. A 24-item questionnaire was developed, with two parallel forms serving as the pre-test and post-test.

Results
Children’s test performance and learning gain are displayed in Figures 2 and 3, respectively.
Children co-creating stories with generative AI showed significant mathematical vocabulary learning gains in the recall dimension (mean difference = 0.61, p = .017) and transfer dimension (mean difference = 0.49, p = .042). However, the learning gains were not significant in the definition dimension (mean difference = -0.22, p = .357) and practice dimension (mean difference = 0.34, p = .305). Looking at the aggregated scores across all four dimensions, while children achieved higher scores on the post-test compared to the pre-test (mean difference = 1.95), this improvement was only marginally significant (p = .058).
In terms of the learning gains by condition, the regression analyses revealed no significant differences between children who co-created stories with generative AI and those who interacted with a present (β = 0.97, p = .300) or hidden human partner (β = 0.76, p = .421). The non-significant condition effect was consistent across all four dimensions of learning.

Scholarly Significance
As child-AI interaction becomes increasingly prevalent in educational contexts, collaborative efforts from educators, AI developers, and psychologists should be made to investigate how children’s learning can be enhanced through such interactions. This investigation will help in understanding children’s developmental trajectories in the digital word and designing child-friendly AI systems that enrich learning experiences.

Authors

• Echo Zexuan Pan, Harvard University

• Xuechen Liu, University of Michigan

• Trisha Thomas, Harvard University

• Chao Zhang, Cornell University

• Ying Xu, Harvard University