How students (really) use ChatGPT: Uncovering experiences among undergraduate students

We examine how undergraduate students integrate ChatGPT into everyday self-directedlearning. We analyze 10,536 naturalistic messages donated by 36 students over morethan a year. A sequential mixed-methods pipeline pairs iterative team-based qualitativecoding with zero-shot language-model annotation validated against human labels (𝜅 =0.75–0.91). It yields a five-category interaction taxonomy of Information Seeking,Content Generation, Language Use, Student–ChatGPT Interaction, and ChatGPTResponse Behavior. Time-lagged linear regression and Cox proportional-hazardsmodels link these categories to sustained engagement. Three findings stand out. First,structured, goal-driven tasks such as theory application, code writing, job-applicationcontent, and multiple-choice questions predict continued use, and ChatGPT becomesincorporated into students’ academic rhythms when gratifications are reliably fulfilled.Second, system-issued apologies are the single strongest positive predictor of increasedengagement, outweighing every task-completion predictor. We name this mechanismrepair gratification, the reward of a breakdown acknowledged and repaired rather than atask simply completed. Third, behaviors marking interactional strain, including promptrevision, frustration, and follow-up clarification, predict disengagement. When the costof managing the system falls on the user without system-side accountability, studentsabandon the tool. We interpret these results through a triangulation of Self-DirectedLearning, Uses and Gratifications Theory, and Expectancy Violations Theory,mapping every predictor onto the four conditions of positive violation,positive confirmation, negative violation, and negative confirmation. Weclose with design recommendations for graduated repair patterns, mode-awareinteraction, and verification affordances, and we outline a participatory AI-literacyagenda for higher education.