A Randomized Study of Human-First and Open Generative AI Instruction for Independent Creative Performance in Sustainability Education

Abstract Generative artificial intelligence (AI) can improve assisted products without establishing independent learning. This three-arm randomized study compared human-first challenger instruction (A), open AI co-creation (B), and a non-AI benchmark (C) in 180 undergraduates at two Chinese universities. Eight workshops accompanied four supervised AI-free assessments. The primary analysis compared post-test creative quality between A and B, adjusting for baseline, section, and assessment form. Among 168 baseline-complete post-test records, A scored 0.41 points higher on the 1–7 scale (95% CI 0.09–0.73; p = 0.013; baseline-standardized Hedges’ g = 0.35). The estimate remained positive with robust standard errors and per-protocol analysis, although two leave-one-section-out intervals included zero. The six-week contrast was 0.41 points (95% CI 0.07–0.75; Holm-adjusted p = 0.112). Secondary creativity, sustainability-reasoning, and evenness comparisons did not pass multiplicity correction. AI-default overlap was exploratory, and workshop breadth and revision counts did not independently predict post-test performance. The findings support a bounded immediate creativity advantage for the combined human-first/challenger package; durable superiority, improved collective diversity, and broader sustainability benefits remain unestablished. Keywords: generative artificial intelligence; human–AI co-creation; sustainable creative learning; independent creativity; conceptual diversity; sustainability reasoning; randomized study