Human-AI Narrative Synthesis to Foster Shared Understanding in Civic Decision-Making

Paper Title

Human-AI Narrative Synthesis to Foster Shared Understanding in Civic Decision-Making

Publication Info

• Topic area: Human-AI collaboration for narrative synthesis in civic engagement.
• Keywords: Human-AI collaboration, narrative synthesis, civic decision-making, large language models, community engagement, storytelling, qualitative data analysis, trust, authenticity, participatory design.

Background and Problem

• Problem / challenge: Traditional methods for synthesizing large-scale community feedback are overwhelmed by the volume of data, especially in contentious civic contexts like school rezoning, limiting shared understanding and empathy across diverse perspectives.
• Significance: Effective synthesis and sharing of community feedback are critical for fostering trust, respect, and informed decision-making in polarized civic environments.
• Motivation and related work: Existing systems focus on raw data presentation or participatory deliberation but lack scalable narrative-based synthesis approaches. AI offers potential but raises concerns about accuracy, transparency, and bias. Narrative storytelling, particularly grounded in lived experiences, has shown promise for fostering empathy and connection but remains underexplored in high-stakes civic contexts.

Solution

• Proposed approach: StoryBuilder, a human-AI collaborative pipeline, transforms large-scale community feedback into first-person composite narratives, deployed via the StorySharer interface for accessible exploration.
• Novelty:
  1. Development of a scalable human-AI narrative synthesis pipeline combining LLM automation with human oversight.
  2. Real-world deployment in a contentious school rezoning process, providing empirical insights into narrative strategies for community engagement.
  3. Controlled experimental evidence on how narrative composition (experience-heavy vs. opinion-heavy) affects interpersonal outcomes like respect and trust.
• Procedure and key techniques:
  • Six-stage StoryBuilder pipeline: data processing, theme creation, theme classification, story generation, human review, and additional theme refinement.
  • Use of McAdams’ Life Story Framework to structure narratives with scenes (experiences) and themes (interpretations).
  • Deployment of 124 composite stories via the StorySharer interface, organized by topic and stakeholder type.
  • Mixed-methods evaluation: field deployment, user studies, and a controlled experiment.

Results

• Concrete findings:
  • Field deployment: 2,183 sessions over four months, with 51% mobile adoption and average session duration of 6.2 minutes.
  • User studies: Participants valued narratives for fostering relatability and understanding but expressed mixed views on AI-authored content.
  • Experiment: Experience-heavy narratives elicited higher respect (3.72) and trust (3.64) than opinion-heavy narratives (respect: 3.29, trust: 3.26).
• Advantage over baselines:
  • Experience-heavy narratives outperformed opinion-heavy ones in fostering respect (+0.43) and trust (+0.38).
  • Mixed narratives showed intermediate benefits but did not outperform experience-heavy stories in trust.
• Experiments / evaluation:
  • Field deployment: Tracked user interactions and feedback on the StorySharer interface.
  • User studies: 21 participants representing diverse community roles provided qualitative insights into narrative impact and system usability.
  • Controlled experiment: 198 participants tested narrative composition effects on respect, trust, and stance change.
• Limitations and future work:
  • Limited validation of all 124 stories before deployment.
  • User study sample skewed toward engaged community members; no student participants.
  • Experiment conducted outside the deployment community, limiting ecological validity.
  • Future work: Improve citation accuracy, explore participatory validation of AI-generated narratives, and integrate quantitative summaries with narratives.

Summary

This study introduces StoryBuilder, a human-AI collaborative pipeline, and StorySharer, an interface for exploring AI-generated narrative summaries of community feedback. Deployed during a contentious school rezoning process, the system synthesized 2,480 community responses into 124 first-person narratives, fostering relatability and understanding across diverse perspectives. Field deployment and user studies highlighted the value of narrative synthesis for community engagement, though concerns about AI-authored content and information overload emerged. Controlled experiments showed that experience-heavy narratives fostered greater respect and trust than opinion-heavy ones but did not shift policy stances. The findings demonstrate the potential of narrative synthesis in civic contexts while underscoring the need for careful design to balance scalability, authenticity, and inclusivity.