The Generation Effect and AI-Assisted Writing: Towards a Framework of Cognitive Fidelity

Abstract

Writing has long been understood as an epistemic activity: a way of developing thought, not merely expressing it. The generation effect in cognitive psychology shows that self-generated material is remembered and understood significantly better than passively received material, and writing research has long argued that the productive struggle of articulation and reformulation is itself a vehicle for knowledge transformation. Generative AI changes the conditions under which this cognitive work occurs. When writers shift from generating to evaluating and selecting AI-produced text, they move from a generation condition to something closer to a reading or response condition, which may lack the cognitive properties that have historically made writing a powerful epistemic tool. This paper argues that critical facets of AI-mediated writing are under-theorized because current research focuses on output quality, efficiency, and integrity rather than on the cognitive value of the writing process itself. We introduce the concept of cognitive fidelity, understood as the degree to which a writing process requires the writer to generate rather than receive the emerging text, as a framework for analyzing these conditions and as a design criterion for evaluating and building AI writing tools. We develop three dimensions of cognitive fidelity (origination, articulation, and reformulation), situate the concept relative to existing frameworks, and draw out implications for the design of AI writing tools and for future research.