Development, Content Validation, and Preliminary Effectiveness of the NEXUS Multicomponent Educational Protocol Integrating AI-Mediated Socratic Inquiry, Identity Meaning-Making and Impactful Expression in Adolescents: A Small-Cluster Randomized Trial with Three-Month Follow-Up
Pages 209-225
https://doi.org/10.5281/zenodo.21862976
Akram Sabzipour, Sajad Hazrati
Abstract Background: Generative artificial intelligence (GenAI) can support explanation and feedback, but unrestricted answers may encourage cognitive offloading, weak source evaluation, and diminished learner authorship. This study developed and content-validated NEXUS, which restricts GenAI to Socratic coaching, and examined preliminary effectiveness for learning agency, intrinsic motivation, and critical-thinking disposition among male upper-secondary students in Qazvin, Iran. Methods: Phase I combined needs mapping in 463 students, co-design, cognitive interviews, and two-round Delphi validation by 15 experts. Phase II randomized 12 classes (106 students; six clusters per arm) to ten weekly NEXUS sessions or a dose- and technology-matched active control. Three-month follow-up was primary. Mixed-model estimates were paired with CR2 inference, CR3 sensitivity analysis, restricted wild-cluster bootstrap-t intervals, and exact within-pair randomization tests. Model-based Hedges g used total class-plus-student variance. Results: The 55-element manual achieved S-CVI/Ave = 0.91; universal agreement was 0.31, and Fleiss kappa was 0.66 for relevance and 0.61 for necessity. Follow-up completion was 89.6%. Adjusted follow-up differences were 0.44 for agentic engagement (95% CI 0.10 to 0.78; Holm-adjusted p = 0.042; g = 0.43), 0.31 for intrinsic motivation (95% CI -0.04 to 0.66; adjusted p = 0.086; g = 0.30), and 6.1 for critical-thinking disposition (95% CI 0.6 to 11.6; adjusted p = 0.078; g = 0.38). The exact global multivariate randomization test was p = 0.031. Fidelity averaged 86.8%; minor AI-output and participant-burden events occurred, but no serious related event was observed. Conclusion: NEXUS showed credible content validity and a promising joint signal, strongest for agentic engagement. With only 12 clusters, incomplete blinding, and no prospective registration, independent replication is required.










