Understanding How Students use ChatGPT in Writing and How Instructors Perceive it

Abstract

The widespread adoption of generative AI in academic writing has created a fundamental asymmetry in education: students make consequential decisions about AI engagement during writing, but those decisions remain invisible to the instructors who evaluate the final product. This dissertation addresses that asymmetry through three interconnected studies grounded in Flower and Hayes' Cognitive Process Theory of Writing, each examining a different dimension of AI-assisted writing at the process level. The first study introduced the NIRVANA dataset and replay platform — a publicly available resource capturing keystroke-level edits, copy-paste events, and ChatGPT interaction histories with temporal granularity. The Human Contribution Ratio (HCR) and Human Edit Ratio (HER) quantify student authorship from process data, and cluster analyses identified four writer profiles: Lead Authors, Collaborators, Drafters, and Vibe Writers. Case analyses demonstrated that surface similarities in final essays can conceal fundamentally different writing processes. The second study observed 77 college students writing argumentative essays with access to ChatGPT, capturing keystroke-level behavior alongside complete ChatGPT interaction histories. We developed a taxonomy categorizing queries into Planning, Translating, Reviewing, and delegation-complete (All) categories, revealing six usage profiles ranging from students who avoided AI entirely to those who delegated the complete writing process — a mode we term Vibe Writing. Writing self-efficacy predicted overall query frequency. Students who used AI for full essay generation reported significantly lower perceived ownership, yet students who used AI selectively for planning or reviewing reported ownership comparable to non-users, even though these uses can bypass the cognitive work writing assignments are designed to develop. The third study examined whether changes in process visibility affect instructor evaluation. Fifty instructors graded six essays under both a static text condition and after observing the writing process through NIRVANA's replay tool. Grades fell from a median of B to C+, with the largest reductions concentrated in essays reflecting heavy AI delegation. The effect was holistic across all rubric dimensions, and instructors who perceived AI as more useful were significantly more likely to flag essays as academic integrity concerns — reflecting AI literacy rather than AI hostility. Together, these studies establish the empirical foundation for process-level accountability in writing education and motivate a prospective system that moves beyond retrospective visibility to embed AI policies directly within the writing environment before students begin.