---
title: 'StorySharer: Civic Narrative Interface'
url: https://www.emergentmind.com/topics/storysharer
type: topic
---

# StorySharer: Civic Narrative Interface

Searching arXiv for the specified StorySharer-related papers to ground the article.
StorySharer is a mobile-friendly interactive web application for presenting AI-assisted, human-reviewed “composite stories” distilled from large volumes of community feedback so that people can explore what a community said during a contentious civic process. In the system described by StoryBuilder, StorySharer is the public-facing dissemination layer rather than the generator itself: StoryBuilder produces the stories, and StorySharer organizes, displays, and contextualizes them for community members and civic leaders. The documented deployment centered on a school rezoning process in a large U.S. school district of roughly 50,000 students, where the platform was used to return thousands of comments to the public in a more readable and navigable form than raw transcripts or dashboard summaries [2509.19643].

## 1. Civic reporting function and conceptual scope

StorySharer was designed for representative democratic settings in which institutions collect large volumes of qualitative feedback but struggle to synthesize and return it in a form that supports public meaning-making. In the reported deployment, the substantive domain was school rezoning, a process described as highly contentious because it implicated school diversity, transportation, access to resources, belonging, safety, and educational opportunity. The system therefore addressed a dual bottleneck: scalable synthesis for officials and accessible interpretation for constituents [2509.19643].

The design choice to present first-person narratives is central. The underlying rationale drew on prior work suggesting that narratives, and especially lived experience rather than abstract opinion, can foster comprehension, empathy, and connection across difference better than expository summaries or lists of opinions. The paper explicitly links this to McAdams’ scenes/themes framework and to literature arguing that personal experiences can bridge moral and political divides more effectively than facts or opinions alone. This suggests that StorySharer should be understood less as a conventional comment browser and more as a civic reporting interface organized around narrative legibility.

A common misconception is that StorySharer is itself the synthesis engine. In the paper, that role belongs to StoryBuilder. StorySharer is the interface through which 124 final stories, derived from 2,480 community “story” and “personal experience” quotes, are presented back to the public [2509.19643].

## 2. Interface organization and interaction model

The interface is organized around a three-tier navigation system. Users first select one of nine topic pages from a left sidebar: Diversity, Transportation, Resources and Programs, School Choice and Housing, Quality and Staff Support, Community Involvement and Trust, Safety and Behavior, Student Well-being, and Change and Disruption. Within each topic, users then select a stakeholder tab—Student Voices, Staff Voices, or Parent Voices—and finally browse the stories within that topic-stakeholder combination. A “More Info” section includes pages such as Project Goals, About This Report, and Story Creation Process [2509.19643].

Each topic page includes explanatory context before the stories themselves: a short topic description, prevalence statistics drawn from the broader community-response corpus, links to project goals and to a separate listening dashboard, and navigation buttons for previous/next topics. Below that, the page presents “Community Experiences,” a note that the stories and themes are based on community feedback from Phase 2 of the rezoning project, and a “share what we missed” button for additional contribution. The resulting structure is report-like rather than feed-like.

Each individual story card contains a stakeholder label or icon, a thematic title, a categorical chip—“Hope,” “Concern,” or “Plus”—and a one-paragraph first-person narrative with inline citation numbers. Hovering on a chip reveals a tooltip definition; clicking inline citations opens an expandable citation section containing original source excerpts, source provenance, and a “report citation” button. Under each narrative, the interface displays the disclaimer: “* This story combines multiple community voices. Click the citations to read their direct quotes.” In the citation accordion, a second disclaimer adds that the story reflects common community feedback on the theme but that not every citation may have been checked, and invites users to report problematic citations with a flag icon [2509.19643].

The interface was designed for accessibility and graduated engagement. StorySharer is mobile responsive, includes multilingual support via a Google Translate widget, and aims for content around average adult U.S. reading level, with themes reduced to 8th-grade level and stories drafted at approximately 5th-grade reading level. A tutorial page shown at entry uses a manually selected example story to explain both project context and mechanics. This suggests a deliberate attempt to make a technically unusual artifact—AI-assisted composite first-person narratives—interpretable without requiring users to infer its epistemic status on their own.

## 3. StoryBuilder pipeline and narrative synthesis mechanics

The StoryBuilder pipeline behind StorySharer has five stages: data processing, theme creation, theme classification, story generation, and automated validation, followed by human review. The source material came from about 8,400 responses from a district-wide online survey and about 170 hours of audio from in-person and virtual facilitated sessions across 119 events, reaching about 3,400 attendees. After processing, the combined material yielded 15,104 distinct quotes totaling about 809k words. For the narrative synthesis system, the authors selected only quotes labeled “story” or “personal experience,” yielding 2,480 quotes totaling about 358k words; 56% of these building blocks came from surveys and the rest from listening-session transcripts [2509.19643].

In data processing, GPT-4o-mini was used to decompose each building block into “scenes” and “themes” using the McAdams Life Story Framework. Scenes captured concrete events and experiences, while themes captured underlying interpretations and values. In theme creation, the authors began from 18 preexisting topics and used both Claude 3.5 Sonnet and GPT-4o to extract detailed themes within each topic and stakeholder type. These models produced 1,048 candidate themes, which a human expert familiar with the data consolidated, using GPT-4o plus manual review, into 190 themes across 9 topics and 3 stakeholder groups. Themes were categorized into pluses, deltas, hopes, and concerns, though delta and concern were later collapsed because the distinction was too subtle; StorySharer ultimately used pluses, hopes, and concerns.

Theme classification mapped each building block to themes using a multi-pass GPT-4o-mini classifier. The paper states that the prompt was applied three times with slight temperature variations and that consensus was computed through set intersection across iterations, requiring agreement across all passes for final theme assignment. In story generation, Claude 3.5 Sonnet selected 3–5 quotes under each theme and “knit” them into first-person narratives, with prompt constraints requiring school-name redaction and a fifth-grade reading level. Each story maintained citation links with unique identifiers mapping back to original community feedback.

A “composite story” is thus a first-person narrative synthesized from multiple community quotes around a common theme and attributed as if spoken by a single first-person narrator, while retaining citations to original excerpts. Earlier in the pipeline there were 190 AI-generated story candidates. After review, 66 were dropped, 44 edited, and 80 left unchanged; the final deployed corpus contained 124 stories. The broader theme inventory contained 301 themes total, of which 177 were theme-only without stories because there was insufficient high-quality narrative material to produce a story [2509.19643].

Automated validation checked five dimensions: minimum three unique citations that are accurately cited, thematic relevance, narrative coherence, stakeholder authenticity, and reading accessibility. Human review then assessed every AI-generated story for relevance, coherence, readability, believability, citation accuracy, and duplication; removed low-quality or redundant stories; made minor edits; manually checked citations on a non-random subset; revised themes for public readability; expanded the theme inventory; and deliberately balanced supportive and critical perspectives across topics and stakeholder groups.

## 4. Evaluation, deployment, and measured effects

StorySharer was evaluated through a four-month field deployment, qualitative user studies with 21 community members, and a preregistered controlled experiment on narrative composition. The field deployment ran from February to May 2025 and was embedded directly into the district’s existing boundary map platform. It was introduced through a district press release and distributed through email, SMS, website banners, social media, physical flyers, and facilitated sessions [2509.19643].

| Evaluation component | Reported evidence | Key result |
|---|---|---|
| Field deployment | 2,183 sessions; 1,675 active for at least 3 seconds | Average time on platform 6.23 minutes (SD = 59.98) |
| User study | 21 community members; 30–60 minute Zoom sessions | Most participants accepted bounded, human-reviewed AI use |
| Controlled experiment | 198 participants after exclusions | Experience-heavy narratives increased respect and trust |

Field analytics show a mixed but informative pattern. Over four months, StorySharer recorded 2,183 sessions, of which 1,675 (77%) stayed active for at least 3 seconds. Users viewed an average of 3.13 +/- 6.97 stories per session. Mobile use was 51%. Only 16 sessions (1%) used the Spanish option. Of active sessions, 194 of 1,675 (11.58%) opened citations. Only 56 sessions (3.3% of active users) submitted story feedback, yielding 233 ratings total. Topic and stakeholder navigation accounted for about half of high-level transitions, but only 550 sessions (32.84%) moved beyond the landing page, and users averaged 2.75 +/- 1.99 topics per session. These data indicate engagement, but also underscore information-overload and onboarding constraints.

The technical evaluation of AI-generated stories was strong on fluency but weaker on provenance fidelity. Exact-match agreement averaged 0.98 for readability, 0.93 for believability, 0.85 for appropriate citation, 0.80 for coherence, 0.82 for relevance, and 0.68 for correct citation. Mean human ratings were reported as readable ($0.99 \pm 0.09$), believable ($0.97 \pm 0.18$), relevant ($0.88 \pm 0.33$), coherent ($0.84 \pm 0.37$), appropriate citation ($0.92 \pm 0.28$), and correct citation only ($0.55 \pm 0.50$). The paper explicitly treats citation accuracy as a significant weakness.

The qualitative study is especially important for interpreting StorySharer’s civic effect. Participants included 13 parents, 4 staff members, 2 parent/staff members, and 2 other community members. Seventeen participants found stories and citations authentic, while eight raised concerns. Twelve participants spontaneously responded to the first story by sharing personal narrative details of their own. The paper interprets this as evidence that the interface did not merely deliver information but often elicited further reflection and storytelling.

The controlled experiment tested four framings of rezoning stories: Scene-dominant (“Experience”), Theme-dominant (“Opinion”), Mixed, and Raw Excerpts (“Control”). Respect showed significant condition effects, $F(3,194)=5.02, p=.002, \eta^2=.072$, with both scene-dominant ($M=3.72$) and mixed ($M=3.77$) narratives exceeding theme-dominant ($M=3.29$). Trust also showed significant effects, $F(3,194)=3.02, p=.031, \eta^2=.045$, with scene-dominant narratives ($M=3.64$) exceeding theme-dominant ($M=3.26$, $p=.038$). Mixed-effects robustness checks likewise found positive coefficients for scene-dominant and mixed narratives on respect and for scene-dominant narratives on trust. No significant effects appeared for stance change or shifts in focus of consideration. The field deployment also found a strong correlation between relatability and trust, $(r = 0.85, N = 81, p\text{-value} < .05)$, while respect and trust in the experiment were correlated at $(r = .78, p < .001)$ [2509.19643].

## 5. Provenance, trust, and contested aspects

StorySharer’s transparency model is explicit but partial. Stories are disclosed as composites, process pages explain how they were created, and citations preserve links back to original source excerpts. Contestability is built into the interface through citation-reporting buttons and a disclaimer acknowledging that not every citation may have been manually checked. This makes provenance visible, but not frictionless: in field use, citations and process explanations were valued in interviews but seldom opened in practice [2509.19643].

One recurrent controversy concerns the synthetic-persona effect. Because multiple voices are composited into a single first-person “I,” some participants found the format more human and less robotic, while others saw it as ethically awkward. One participant described it as an “avatar parent.” The paper presents this as a genuine design tension rather than a solved problem: compositing protects individual identity and can increase relatability, but it may also blur the boundary between faithful synthesis and invented persona.

A second controversy concerns AI authorship in high-stakes civic settings. Most participants in the real deployment accepted AI use when it was bounded and human-reviewed, especially given the scale of the data. Five participants nevertheless voiced critiques, including distrust of AI for something “this important.” The paper argues that acceptance was context-sensitive: participants inside the live district process often saw AI as a practical aid for handling an “obscene amount of information,” whereas experiment participants outside that context more readily interpreted AI-authored narratives as potentially deceptive or manipulative.

Privacy and authenticity were handled through redaction, aggregation, and citation-based accountability rather than formal privacy guarantees. School names were redacted in generation prompts. Composite storytelling reduced one-to-one testimonial exposure. Source attribution and process explanations supported authenticity claims. At the same time, the authors explicitly note unresolved risks around representation, subtle distortion, and citation errors. They recommend strong human validation at critical points and suggest participatory validation by original contributors as future work. A plausible implication is that StorySharer should be seen as a hybrid civic-reporting instrument whose legitimacy depends as much on governance and review procedures as on model quality.

## 6. Related systems and design lineages

StorySharer belongs to a broader family of systems that combine story representation, sharing, and computational mediation, but it occupies a distinctive position within that family. Web-STAR is a browser-based platform for symbolic story understanding built on the STAR reasoning engine, with a public repository, commenting, collaborative editing, and web services; it treats stories as formal, executable, and discussable objects rather than as narrative reports for a general public [1808.00048]. SHARI, by contrast, turns the “biggest story of the day” into a visual historical narrative by combining StoryGraph, Hypercane, ArchiveNow, and Raintale; it is closer to automated retrospective news storytelling than to civic feedback synthesis [2008.00139].

Other systems illuminate adjacent dimensions of the design space. A web-based tool for social-video repurposing automates one-click generation of multiple platform-ready summaries from a full-length video by combining AC-SUM-GAN-based video summarization with saliency-driven aspect-ratio transformation; this addresses dissemination-format adaptation rather than narrative synthesis from public testimony [2312.02616]. “Tell Your Story” frames montage creation over personal media as a task-oriented dialog problem, using the C3 dataset to model CREATE, REMOVE, REPLACE, REORDER, REFINE, and MODIFY_DURATION operations over an evolving montage state; this is closely aligned with interactive content creation, but its output is a media montage rather than a cited civic narrative report [2211.03940]. TaleCrafter provides a modular pipeline for visualizing stories with multiple characters through story-to-prompt generation, text-to-layout generation, controllable text-to-image synthesis, and image-to-video animation; it is oriented toward story visualization and interactive editing rather than public voice synthesis [2305.18247]. SceneAR extends short-form storytelling into augmented reality through sequential scene-based micro narratives that can be shared and remixed as AR content, emphasizing in situ authoring and spatial remix rather than civic reporting [2108.12661].

Taken together, these systems suggest three recurrent design axes around StorySharer. One axis concerns whether stories are primarily for reasoning, dissemination, or public interpretation. Another concerns whether sharing means publishing static outputs, editable artifacts, or socially remixable experiences. A third concerns whether the system returns source material directly, summarizes it, or synthesizes new narrative forms from it. StorySharer is unusual because it combines large-scale qualitative synthesis, first-person composite narration, explicit provenance cues, and deployment inside a live representative decision process.

Source: https://www.emergentmind.com/topics/storysharer