Enhancing Peer Exposure and Creating Opportunities for Bridging Social Capital Through an AI-Generated Summary-Driven Learning Design in Online Discussion Forums

Synthesis: This design-based research study by Hao and Cukurova evaluates an AI-generated summary-driven learning design (AI-SLD) in online collaborative discussion forums with 128 university students across three design iterations (baseline, Wizard-of-Oz, and full AI). Using social network analysis of viewing logs and thematic analysis of interviews, it shows that AI-generated discussion summaries significantly broadened students' exposure to peer contributions and strengthened network connectedness, functioning as navigational scaffolds that lowered the effort of locating meaningful posts and expanded opportunities for building the weak-tie connections social-capital theory terms bridging social capital. However, the support did not prevent a decline in viewing activity under rising academic workload, and the effect was moderated by discussion topic and topic familiarity.

Summary

Prior research on online discussion engagement has largely targeted posting and replying while overlooking the viewing activities that precede and enable them, and has focused on human-written rather than AI-generated summaries. This paper proposes and tests an AI-generated summary-driven learning design (AI-SLD) that uses Large Language Models (LLMs)-produced discussion summaries and example posts to lower the search cost of navigating large, poorly-titled forums and thereby broaden students' viewing breadth and access to heterogeneous peer perspectives (an opportunity condition for bridging social capital). Conducted via design-based research across three iterations with 128 higher-education students over an eight-week intervention, the study combines quantitative social network analysis of forum log data with qualitative thematic analysis of student interviews.

Key Findings

• AI-SLD broadened viewing breadth. Students in the AI-supported iterations (PSG, MSG) showed significantly higher standardized out-degree centrality than the baseline group (β = 0.547 and β = 0.438, both p < 0.001), indicating they viewed/interacted with more peers; the Wizard-of-Oz and full-AI versions performed comparably, showing AI-generated summaries can serve as a scalable alternative to human-written ones.
• Summaries acted as navigational scaffolds, not mere content. Interview data described the AI summaries as a "navigation map": rather than increasing the sheer number of posts viewed, they reduced the cognitive and temporal effort of finding relevant or interesting contributions buried in ill-named threads — a finding consistent with framing them as scaffolds rather than content substitutes.
• Viewing served selective attention and epistemic diversity. Students deliberately attended to posts from visible "active contributors," drawn not by friendship but by contrasting opinions and different backgrounds, suggesting viewing breadth may carry epistemic value relevant to forming the bridging (weak-tie) form of social capital.
• Support was moderated by topic familiarity and academic workload. When topics felt unfamiliar, students consulted more peers; familiar topics reduced engagement breadth. The AI-SLD did not prevent declining viewing activity in later weeks — a trend students attributed to draft-assignment deadlines and shifting priorities rather than dissatisfaction with the design.
• Network analysis distinguished peer-to-peer from AI-mediated viewing. Only the full-AI iteration (MSG) showed significantly higher raw-network density (p = 0.045) than baseline once AI-generated nodes were included, indicating the summaries and example posts contributed to the overall viewing network beyond direct peer interactions.

What this means for practice

• Instructors. Post an AI-generated summary of the week's discussion, plus one AI example post, as a navigation map before students read threads; the supported iterations raised viewing breadth (standardized out-degree centrality β = 0.547 and 0.438 versus baseline, both p < 0.001).
• Instructors. Do not rely on the summaries to hold participation: viewing activity declined in later weeks across all iterations and students attributed it to draft-assignment deadlines, so pair the Scaffolding with socio-pedagogical strategies.
• Instructors. Deploy the design when topics are unfamiliar, since students consulted more peers when a topic felt new and narrowed their viewing when it felt familiar.
• Researchers. Connect viewing behavior to discussion quality and cognitive engagement, since log-based Learning Analytics cannot show whether viewing was deep or shallow, and treat viewing and attentional behavior as observable evidence of students' opportunities to access diverse peer perspectives — the condition for developing bridging social capital.

Limitations

• The three iterations are separate cohorts in different academic years (baseline n = 49 in 2021–2022, pilot n = 40 in 2022–2023, main n = 39 in 2023–2024), so cohort-level differences such as post-pandemic online-learning experience may explain the behavioral shifts, and the findings are not causal.
• Social network analysis of Moodle log data cannot determine the depth or quality of learners' viewing engagement.
• The planned random-slopes model failed to converge because many students did not participate in all 8 weeks, forcing a random-intercepts-only model, and OLS replaced mixed effects for density because only three groups existed.
• Interview evidence is partial: 31 students from the pilot and main cohorts only (about 60 minutes each, recruited voluntarily after grades were released), and pilot interviewees were told the summaries were human-written.

Citation

Hao, X., & Cukurova, M. (2026). Enhancing peer exposure and creating opportunities for bridging social capital through an AI-generated summary-driven learning design in online discussion forums. Journal of Computer Assisted Learning, 42(5), e70317.