An Experimental Study Exploring Human–AI Complementarity in Early Social-Emotional Learning

Synthesis: Raave, Colasante, Roldan Roa, Ramos Martinez, Li, and colleagues (2026) run a controlled comparison of a Generative AI pedagogical conversational agent (PCA) and human educators independently facilitating story-based social-emotional learning activities with a static AI child, using 18 simulations per facilitator across three activities (108 observations) rated by expert coders blind to facilitator type. The PCA showed strengths in basic relational and instructional domains — maintaining a respectful tone and routinely providing procedural Scaffolding — while educators showed strengths in deeper SEL instruction, guiding Reflection and promoting social-emotional knowledge.

Key Findings

• Expert raters blind to facilitator type found significant differences between the PCA and human educators on dialogue indicators.
• The PCA excelled at basic relational and instructional domains: maintaining a respectful tone and providing routine procedural scaffolding.
• Human educators excelled at deeper SEL instruction: guiding reflection and promoting social-emotional knowledge.
• Findings support a human–AI complementarity framing rather than treating PCAs as direct substitutes for educators in SEL.

What this means for practice

• Instructors. Delegate routine procedural scaffolding and relational steadiness to a pedagogical conversational agent and reserve your own time for guiding reflection and promoting social-emotional knowledge, where human educators outperformed the PCA.
• Instructors. Plan a division of labor rather than a substitution: the PCA was more consistent in rapport and emotional tone and broke SEL tasks into discrete steps more reliably than educators did.
• Instructors. Anticipate that human facilitation varies: the educators' lower consistency in rapport and emotional tone may reflect fluctuations in energy, mood, or interpretation of the AI child's cues, which matters for emotional safety in early childhood settings.
• Researchers. Replicate in vivo, since a static AI child with scripted, voice-based interactions cannot capture how educators adapt to real children's nonverbal cues, cultural backgrounds, and developmental needs.
• Researchers. Add Multimodal AI evaluation — the solely text-based coding cannot detect differences in warmth, authenticity, or cultural responsiveness conveyed through tone and gesture.

Limitations

• Facilitators were 18 educators (ages 24–40, M = 30.3 years; n = 16 female) recruited through convenience sampling across international early childhood and primary settings, compared with 18 PCA simulations, yielding N = 108 coded observations across three activities.
• The primary stated limitation is reliance on a static AI child with voice-based interactions rather than real children in vivo, which the authors say limits ecological validity and may underestimate educators' performance relative to the PCA.
• The analysis was solely text-based, so subtle pedagogical and emotional nuances that facilitators convey through tone and gesture were unavailable to the expert raters.
• Educators' expected advantage on growth-oriented nudging is likely larger with known students, so the present comparison should be read accordingly.

Citation

Raave, D. K., Colasante, T., Roldan Roa, E., Ramos Martinez, J. C., Li, H., et al. (2026). An experimental study exploring human–AI complementarity in early social-emotional learning. Computers and Education Open, 10, 100331.