Same AI, Different Papers: Measuring Variance in AI-assisted Student Writing

Background: Debates about generative AI in higher education claim that allowing unlimited AI access compresses academic writing quality such that student contribution becomes marginal. The output-compression hypothesis predicts manuscripts should converge toward a narrow quality band when students share the same capable AI tool.Objectives: This study examines whether unlimited generative AI access in an undergraduate qualitative methods course produces compression or maintains meaningful variance in student writing quality, and traces plausible pathways that may account for divergent outcomes.Methods: The study analyzes variance in final research papers using rubric-proximal proxy indicators (claim-evidence coupling, theory-interpretation linkage, methods specificity, scholarly sourcing) and examines six contrasting cases through longitudinal artifacts and AI interaction logs.Results: Results show substantial dispersion in claim-evidence coupling, theory-interpretation linkage, and scholarly sourcing. One dimension shows compression: methods specification language, where AI assistance combined with scaffolding narrows gaps. Boundary condition suggests compression occurs where demands are structural rather than judgmental. Case analyses reveal stronger outcomes correlate with distinct orchestration practices including precise task specification, irreducible inputs provision, iterative revision, and critical evaluation. Patterns align with Extended Executive Cognition.Conclusions: Findings establish that AI-enhanced pedagogy remains viable where variance persists on judgment-requiring dimensions. Student orchestration of AI-supported workflows determines quality when structural demands are held constant.