Design-thinking scaffolds for learner-agency-preserving AI use: effects on creative self-efficacy, cognitive flexibility, and engagement

Introduction Generative AI can facilitate the production of student work but may also displace learners' responsibility for problem framing, evaluation, and revision. This study examined whether design-thinking scaffolds were associated with creative self-efficacy, cognitive flexibility, and problem-solving engagement during AI-supported project learning. Methods A three-condition, four-wave quasi-experimental study included 162 undergraduates in six intact class sections (54 per condition). Non-AI project learning, ordinary GenAI use, and design-thinking-scaffolded GenAI use were compared across a four-week educational design task. Baseline-adjusted models assessed self-reported outcomes, human-coded AI-use logs, and independently rated products from 54 teams. Cluster-bootstrap analyses examined temporally ordered indirect associations and moderation by pre-intervention AI over-reliance. Results Compared with ordinary GenAI, the scaffolded condition showed higher T2 creative self-efficacy (adjusted difference = 0.33, p = 0.006), T3 cognitive flexibility (0.35, p = 0.003), and T4 engagement (0.35, p = 0.004). Mean team product quality was 20.14 in the scaffolded condition, compared with 17.62 under ordinary GenAI and 16.76 without AI. Coded prompts showed fewer answer-completion requests and more comparison, critique, and reflection. Indirect estimates via creative self-efficacy (0.07), cognitive flexibility (0.09), and their sequential pathway (0.04) had bootstrap confidence intervals excluding zero. Higher initial AI over-reliance weakened the overall engagement contrast and the two single-mediator indirect pathways, whereas the moderated-sequential index had a confidence interval including zero. Discussion The findings are consistent with agency-preserving cognitive offloading when instructional scaffolds retain student responsibility for higher-order decisions. Intact-section allocation, task-specific self-reports, and limited section-level replication preclude strong causal or general-transfer claims.