Abstract
This study investigates the distinct mechanisms of human versus Large Language Model (LLM) creativity. Employing a two-stage experimental design, we systematically compared Human-Only, LLM-Only, and LLM-Assisted performance across propositional and creative writing tasks. Results revealed a critical asymmetry contingent upon the research context: human authors exhibited higher originality in high-demand creative tasks, whereas LLMs governed execution quality, maintaining superior effectiveness across different tasks and cohorts. This pattern is characterized by four exploratory writing creativity profiles: Ideal, Safe, Moderate, and Plain. The distribution of human and LLM writings across these profiles was strikingly different. Hierarchical Moderated Regression analysis uncovered divergent linguistic pathways: human originality is predicted by markers of subjective cognitive investment, while LLM effectiveness is mechanistically driven by optimized structural coherence. Furthermore, the study identified a "Collaboration Trap" during collaboration with a suboptimal LLM. This partnership failed to improve human performance relative to LLM-Only benchmarks and induced cognitive anchoring, leading humans to mimic AI complexity without quality gains. These insights offer critical implications for preserving human agency and avoiding homogenization in future human-AI collaborative writing pedagogies.