Fabula: Building a Narrative Storytelling Sidekick with the Writers' Community

This paper presents Fabula, an interactive AI-powered app for fiction writers that utilizes hierarchical narrative planning and participatory design with 42 experts to evaluate how generative models, transparent narrative structures, and iterative planning tools impact story quality and writer creativity across diverse storytelling traditions.

The Gist

Imagine you are trying to build a house. You have two ways to do it:

1. The Architect: You draw a perfect, detailed blueprint first, room by room, before laying a single brick.
2. The Gardener: You dig a hole, plant a seed, water it, and see what grows, adjusting as you go.

For a long time, AI writing tools were like a chatty friend who just talked at you, giving you random paragraphs whenever you asked. The team at Google DeepMind wanted to build something different. They created Fabula, a digital assistant for writers that tries to support both the Architect and the Gardener.

Here is the story of how they built it, what they learned, and why they had to rethink their original plan.

The Big Idea: A Story "Control Panel"

Fabula isn't just a chatbot. It's a workspace that shows you two things at once:

• The Script: The actual story text (the dialogue and action).
• The Plan: A detailed map of the story's structure, broken down into "Scenes" (big chunks) and "Beats" (small moments of change).

Think of the Plan as the skeleton of the story. The Script is the skin and muscles. Fabula lets you tweak the skeleton, and the skin adjusts automatically. Or, you can write a line of dialogue, and the skeleton updates to match.

The Experiment: Building with Writers, Not Just for Them

The team didn't just build the tool in a lab and hand it to writers. They treated the tool like a "cultural probe"—a way to poke and prod the writing community to see what actually works. They invited 42 experts (screenwriters, playwrights, professors) and hundreds of others to play with it.

Here are the four big lessons they learned, explained simply:

1. The "Auto-Editor" Problem: Perfect vs. Surprising

The Team's Guess: They built an AI "judge" (an auto-rater) that checked stories for consistency, clear goals, and logical endings. They thought, "If the AI makes the story consistent, it will be better."
The Writers' Reality: Writers loved the structure but hated the predictability. They told the team, "Great stories need to be weird, surprising, and sometimes break the rules." The AI was too good at making things "safe" and "logical," which killed the magic.
The Fix: They realized the AI shouldn't just try to be perfect; it needs to allow for "absurdity" and unexpected twists, even if they seem messy.

2. The Interface: Too Many Buttons?

The Team's Guess: They built a complex three-column interface (The Architect view) to show the whole story plan. They thought writers would love having total control over the blueprint.
The Writers' Reality: Many writers found it overwhelming and rigid. It felt like trying to drive a car with a manual transmission when you just wanted to cruise. Some felt the tool forced them to plan too much before they could write.
The Fix: They added a "Lock" button (so you can freeze parts of the story so they don't change when you edit elsewhere) and a "Chatbot" sidekick for those who just want to talk to the AI without looking at the complex charts.

3. The Cultural Bias: The "Western" Story Spine

The Team's Guess: They used classic storytelling rules (like Aristotle's three-act structure) to build the story plan. They thought this was a universal language for stories.
The Writers' Reality: Writers from different cultures pointed out that this felt very "Western." It forced stories into a specific shape (beginning, middle, end, character growth) that doesn't fit every culture's way of telling tales. The AI also kept defaulting to white, middle-class characters even when writers tried to be specific.
The Fix: They acknowledged that one size doesn't fit all. They are working on making the tool more flexible so it doesn't force every story into a Western mold.

4. The "Convergent Iteration" Debate: Finding vs. Creating

The Team's Guess: They viewed writing as a "search" process. They thought, "If we generate 100 versions of a scene, the writer can just pick the best one." They called this "Convergent Iteration."
The Writers' Reality: This sparked a huge debate.

• Novice writers loved it. It helped them overcome the fear of the "blank page."
• Pro writers were worried. They felt that if the AI does the heavy lifting, they might lose their "writing muscles" (cognitive deskilling). They worried about becoming just "editors" of machine text rather than creators.
  The Fix: They shifted the tool's focus. Instead of just generating whole stories to replace the writer, they are turning Fabula into a tutor. It's now better at analyzing a writer's existing draft and asking, "Hey, what if this character had a different motivation?" rather than just writing the scene for them.

The Final Takeaway

Fabula is not a robot that writes movies for you. It is a collaborative playground.

The team realized that the biggest tension isn't about technology; it's about control. Writers want to feel like they are the masters of their stories, not just passengers in an AI car.

By listening to the writers, the team is now pivoting Fabula away from being a "story generator" and toward being a creative sidekick. It's becoming a tool to help students learn structure, to help screenwriters brainstorm quickly, and to help theater groups improvise live stories.

In short: The team built a high-tech car, but the drivers (the writers) told them, "We don't want an autopilot; we want a really good co-pilot who knows the map but lets us steer."

Technical Summary

Technical Summary: Fabula – Building a Narrative Storytelling Sidekick with the Writers' Community

Problem Statement

The integration of Large Language Models (LLMs) into creative writing presents a tension between the potential for transformative interactive storytelling and significant concerns regarding authorship, agency, and the homogenization of cultural narratives. While AI tools offer new forms of co-creation, the writers' community is fractured: some embrace LLM-based co-writing, while others reject them due to ethical concerns or fears of "creative deskilling." Existing AI writing tools often rely on chat-based interfaces that lack structural depth, or they fail to accommodate diverse writing archetypes (e.g., "Gardeners" vs. "Architects") and non-Western storytelling traditions. Furthermore, there is a lack of participatory frameworks that treat user feedback not merely as a mechanism for tool optimization ("participation washing") but as a method to negotiate fundamental values and ideological conflicts between developers and artists.

Methodology

The authors present Fabula, an interactive web application designed to support fiction writers through a hierarchical narrative planning system. The development and evaluation of Fabula followed a participatory design approach, utilizing the tool as a "cultural probe" and "boundary object" in adversarial design to surface conflicts between technical affordances and artistic needs.

System Architecture

• Drama Managers: The core engine consists of modular Python interfaces (Drama Managers) that orchestrate story generation. These managers operate on three levels of abstraction: the Story Plan (high-level arc), the Beat Plan (scene-level structure), and the Script (dialogue and action).
• Hierarchical Modeling: Stories are structured into scenes and beats, drawing on classical narratology (e.g., McKee, Stanislavsky). The system enforces a "Drama Manager" contract that includes:
  • Plan Creation: Generating initial hierarchical plans.
  • Content Generation: Producing script lines for beats.
  • Plan Evolution: Dynamically updating plans based on user feedback.
• Self-Improving Agents: The system employs a "self-play" loop where agents generate stories, which are then evaluated by an Auto-Rater. The Auto-Rater uses an LLM (Gemini 2.0/2.5) to score narratives against 63 guidelines distilled from narrative theory and creative writing textbooks (e.g., Story, Into the Woods).
• User Interface (UI): Fabula offers two distinct views:
  • Gardener View: Focuses on the script and immediate scene generation (improvisational).
  • Architect View: Displays the story plan alongside the script, allowing for hierarchical editing and structural oversight.

Evaluation Protocol

The study involved a multi-phase iterative development cycle:

1. Internal Testing: Hundreds of colleagues and internal experts.
2. Design Interviews: Qualitative sessions with 18 external experts (academics, playwrights, screenwriters, TV executives) recruited via snowball sampling.
3. Writer Sessions: 25 external writers participated in 3-hour mixed-method sessions involving tool interaction, creativity surveys (Creativity Support Index), and focus groups.
4. Large-Scale Testing: A limited public release to over 250 writers (internal and external testers) for quantitative feedback on usability and inspiration.

Key Contributions

1. A Hierarchical Narrative Interface: Fabula departs from standard chatbot interfaces by providing a dual-view system (Gardener/Architect) that supports both granular, moment-by-moment creation and high-level structural planning.
2. Participatory "Adversarial" Design: The paper demonstrates a methodology where the tool itself is used to interrogate design choices. Rather than simply fixing bugs, the team used feedback to expose ideological clashes (e.g., the tension between algorithmic consistency and the artistic value of "absurdity" or "de-familiarization").
3. Auto-Rater Validation: The authors developed and validated an LLM-based auto-evaluator that correlates significantly with human preferences (0.83 correlation) for story quality, providing a scalable method for iterative improvement.
4. Cultural and Structural Probing: The study critically examines the applicability of Western-centric narratological models (e.g., three-act structure) to diverse cultural traditions and writing mediums (theatre vs. screenwriting), identifying specific biases in LLM outputs regarding character representation and cultural nuance.

Results and Findings

• Story Quality vs. Style: The system excelled at mechanical aspects of storytelling (structure, scene breaking, and plot coherence) but struggled with stylistic nuance, irony, and specific authorial voices. Writers found the output "generic" or "mainstream," lacking the intentional rule-breaking characteristic of specific literary styles.
• UI Preferences: While the Architect view was praised for its scaffolding and ability to restructure narratives, many writers found it "rigid" and cognitively overwhelming. Conversely, the Gardener view was appreciated for its granular flow but sometimes felt like a "computer game." A significant finding was the desire for a hybrid approach, leading to the addition of a collapsible chat interface alongside the structured views.
• Cultural Bias: LLMs demonstrated a tendency to default to Western, hegemonic social groups (e.g., white, middle-class characters) even when prompts were neutral. Writers noted that non-Western storytelling structures were often treated as superficial embellishments on a Western narrative spine.
• Convergent Iteration: The "search and discovery" paradigm (finding a story rather than creating it from scratch) was highly valued by novice writers and screenwriters for ideation and world-building. However, theatre practitioners and experienced writers expressed concerns about cognitive deskilling, fearing that reliance on AI might erode the "thinking" process inherent in writing.
• Authorship and Agency: Writers expressed mixed feelings about ownership. While some felt agency through extensive editing, others feared the "ghostwriting" effect and the loss of the unique "authorial voice." There was a strong consensus that the tool should act as a "sidekick" or "diagnostic partner" rather than a replacement for the writer.

Significance and Claims

The paper claims that Fabula serves as a critical design probe that reveals the boundaries of what writers are willing to cede to AI. Its primary significance lies not in delivering a perfected commercial product, but in:

1. Reframing the Role of AI: Shifting the paradigm from "story generation" to "analytical and diagnostic support," positioning the AI as a tutor or a "sounding board" for structure and character development.
2. Highlighting Ideological Conflicts: Demonstrating that participatory design in AI should not aim to resolve all tensions but to make conflicting logics (e.g., algorithmic optimization vs. artistic friction) visible and subject to debate.
3. Educational Potential: Identifying a strong use case for Fabula as a pedagogical tool for teaching narrative structure to novices and for interactive theatre devising, where rapid iteration and structural analysis are more valuable than final text generation.

The authors conclude that the development process itself is a form of "convergent iteration," allowing the team to identify limitations in both their assumptions and the underlying technology, ultimately guiding the tool toward specialized applications in education and interactive storytelling rather than mass content generation.