How Students (Really) Use ChatGPT: Uncovering Experiences Among Undergraduate Students

TL;DR

The paper investigates how undergraduates actually use ChatGPT in self-directed learning, using naturalistic interaction logs from 36 students and drawing on Self-Directed Learning (SDL) and Uses and Gratifications Theory (UGT). Through a three-phase mixed-methods design (qualitative coding, GPT-4o scaling labeling, and explanatory analysis) plus temporal analyses, the study identifies five broad use categories and establishes behavioral predictors of engagement and retention, including the pivotal role of conversational repair and apologies. The findings reveal that task-focused interactions (e.g., theory application, coding, job applications) and effective repair dynamics sustain long-term use, while persistent breakdowns or misalignment can decrease engagement. The authors translate these insights into design and policy recommendations—emphasizing repair scaffolding, adaptive mode transitions, epistemic vigilance, and participatory AI literacy—to help educators integrate generative AI in higher education in a transparent, pedagogically grounded manner.

Abstract

This study investigates how undergraduate students engage with ChatGPT in self-directed learning contexts. Analyzing naturalistic interaction logs, we identify five dominant use categories of ChatGPT: information seeking, content generation, language refinement, metacognitive engagement, and conversational repair. Behavioral modeling reveals that structured, goal-driven tasks like coding, multiple-choice solving, and job application writing are strong predictors of continued use. Drawing on Self-Directed Learning (SDL) and the Uses and Gratifications Theory (UGT), we show how students actively manage ChatGPT's affordances and limitations through prompt adaptation, follow-ups, and emotional regulation. Rather than disengaging after breakdowns, students often persist through clarification and repair, treating the assistant as both tool and learning partner. We also offer design and policy recommendations to support transparent, responsive, and pedagogically grounded integration of generative AI in higher education.

Figures (13)

• Figure 1: Timeseries of the conversations in our dataset. We can clearly see decreased activity during Spring break (March 9-17) and Summer 2023 indicating that most usage was academic.
• Figure 2: Sequential mixed-methods integrating human qualitative expertise with computational scaling. Human researchers developed a grounded theory codebook (Phase I), which structured GPT-4o annotation across 10,536 messages (Phase II). Validation confirmed strong human-machine agreement (Cohen's $\kappa$ = 0.75--0.91), enabling temporal analyses linking interaction patterns to engagement trajectories (Phase III).
• Figure 3: Breakdown of log interactions based on the five top categories.
• Figure 4: Subcategories within Information Seeking prompts, including concept explanation, theory application, and clarification of instructions.
• Figure 5: Subcategories within Content Generation, such as code writing, job application content, multiple choice solving, and summarization.