A recent experimental study by Raave and colleagues employed a mixed factorial design to compare a pedagogical conversational agent (PCA) with human educators in facilitating early social-emotional learning (SEL). Eighteen international early childhood and primary teachers and an AI-driven PCA each conducted three standardized story-based SEL activities with a static AI child named “Ali,” designed to exhibit challenging characteristics including distractibility and limited empathy. This controlled comparison yielded 108 total observations across activities focusing on prosocial behavior, perspective-taking, and healthy guilt induction.
Expert raters, blind to facilitator type, coded interactions for three evidence-based pedagogical support techniques—cognitive-conceptual, procedural, and emotional support—using binary scales, and assessed broader teaching quality across six dimensions adapted from the Danielson Framework for Teaching on a 1-4 scale. The study maintained high inter-rater reliability across all measures (Cohen’s κ=.65−.80 for support techniques; quadratic weighted κ=.61−.76 for quality indicators).
Results revealed significant differences in facilitation approaches. Mixed ANOVA showed educators used more support techniques overall (M=7.57 vs. M=5.47) and substantially more cognitive-conceptual support than the PCA (M=15.28 vs. M=6.46, p<.001), while the PCA provided slightly more procedural support (p=.026). Quality ratings demonstrated complementary strengths: the PCA scored higher in relational domains such as respect and rapport (d=1.86) and flexibility (d=1.06), whereas educators excelled in intellectual engagement including questioning techniques (d=0.72) and engaging students in learning (d=0.96).
The findings reveal meaningful pedagogical asymmetry suggesting human-AI complementarity rather than competition. The PCA maintained consistent emotional validation and procedural structure but tended to be “caring but pedagogically shallow,” while educators provided deeper cognitive-conceptual scaffolding and moral reasoning guidance with greater affective variability. The authors propose a strategic division of labor where PCAs manage routine structuring and baseline emotional support, while human educators focus on complex social-emotional reasoning and ethical development. Future research should explore direct co-facilitation models in authentic classroom settings.
Source (Open Access): Raave, D. K., Colasante, T., Roa, E. R., Martinez, J. C. R., Li, H., Mukherjee, S., & Malti, T. (2026). An experimental study exploring human–AI complementarity in early social-emotional learning. Computers and Education Open, 100331.