GenAI in Online Collaborative Argumentation: Comparing Interaction Patterns between High and Low Performers

Research is exploring how generative AI (GenAI) can support online collaborative argumentation (OCA), yet little is known about how student characteristics (e.g., domain knowledge) influence their interactions with GenAI and argumentation quality. Examining students-GenAI interactions (SAI) could aid in unpacking the underlying GenAI-empowered OCA mechanisms to open the black box of OCA processes to facilitate learning behavior. To this end, we collected chat histories of 60 graduate students and their final argumentative essays through a ChatGPT-powered OCA system on Discord. We employed epistemic network analysis to reveal SAI patterns during OCA. We conducted the Wilcoxon rank-sum (Mann-Whitney U) test to evaluate the OCA performance differences between different levels of domain knowledge. Findings revealed that high domain knowledge (HD) students engaged in more multifaceted OCA activities, demonstrating higher quality argumentation than low domain knowledge (LD) students. The HD group assessed GenAI's inputs more constructively, synthesizing human and AI contributions. Additionally, GenAI demonstrated a higher frequency of performing phases of OCA when interacting with HD students. Significant differences in argumentation quality across various evaluation criteria were confirmed. This study offers implications for improving the design and implementation of GenAI in OCA.