PaperlayerView exact revision on arXiv

Source-linked AI summary

The House with a Million Windows: Interactive Fiction for Narrative Restorying

Cody Kommers, Sarah G Immel, Drew Hemment, Mina Lee

arXiv:2609.12537v1cs.CLcs.HC

TL;DR

AI-assisted writing raises concerns about flattened meaning and questions about its effects on narrative practices and personal identity. HWAMW guides users through retelling personal stories with LLM-generated variations. Studies report shifts in users’ narratives and narrative identity, while the current prototype focuses on intrapersonal meaning.

  • Problem

    The paper asks how AI-assisted writing may affect human narrative practices and individuals’ sense of identity and meaning, amid concerns about flattened meaning and displaced intentional storytelling.

    Source: §1 · p. 2

  • Method

    HWAMW is an interactive restorying prototype in which users tell a personal story, reflect on LLM-generated variations, and retell it.

    Source: §1 · p. 2

  • Results

    HWAMW users’ narratives and narrative identities shifted, and the authors report a significant increase in users’ sense of narrative identity.

    Source: §1 · p. 3, §6 · p. 10

  • Takeaways & Limitations

    HWAMW suggests a way to position LLMs as counternarratives that help people encounter and reconsider their own stories rather than replace their storytelling.

    Source: §6 · p. 10

    The current prototype targets intrapersonal meaning rather than cultural meaning or stories’ real-world impacts.

    Source: §7 · p. 11

Abstract

from arXiv · show

AI-assisted writing can flatten meaning in human storytelling, enabling the production of homogeneous outputs without the intentional effort and sense-making writing entails. To address this challenge, we present The House with a Million Windows (HWAMW), an LLM-based interactive fiction system designed to help users explore both the breadth and depth of potential meanings within their personal stories -- drawing on a psychological paradigm called the restorying intervention. In HWAMW, users play through a text-based narrative in which they tell a story, then encounter a set of LLM-generated "windows" reframing it according to different literary styles. Empirical evidence shows that HWAMW increases users' sense of narrative identity, while an expert review explores how this effect is achieved. Our findings suggest that HWAMW facilitates restorying and offers a valuable paradigm for AI-assisted writing, wherein LLMs do not tell our stories but rather help us see greater potential in the stories we tell.

1 Introduction

AI-assisted writing raises concerns about flattened meaning and displaced human storytelling, as well as questions about narrative identity. HWAMW uses LLM-generated variations within a restorying experience and evaluates the prototype with general users and expert reviewers.

1 Introduction

  • The paper addresses the risk that AI-assisted writing can flatten meaning by displacing the effortful process through which people think through personal experiences.The authors also note that AI-assisted outputs tend toward cultural or stylistic homogeneity, while treating this risk as a consequence of deployment rather than an intrinsic property of the technology.Source: §1 · p. 2
  • HWAMW guides users to tell a personal story, reflect on LLM-generated variations, and retell it, aiming to preserve their authorship and agency.Its variations serve as possible perspectives on the user’s story rather than drafts to accept.Source: §1 · pp. 2–3
  • HWAMW brings together narrative and literature, AI-driven interactive writing, and psychological restorying interventions.The paper positions this combination at a novel intersection of the three research strands.Source: §1 · p. 2
  • HWAMW is a reflective writing interface, not a clinical therapeutic aid, and therefore does not provide therapy’s licensed-practitioner expertise and oversight.The authors describe narrative-driven restorying as both more limited and more general than therapeutic work.Source: §1 · p. 2
  • The design aims to support exploratory restorying, use literary fiction to expand available narrative techniques, and avoid implying clinical effects.The authors frame the experience more like a creative writing workshop or literary adaptation than a therapeutic aid.Source: §1 · pp. 2–3
  • A crowdsourced study (N=40) found participants’ narratives and narrative identities shifted during HWAMW; an expert review (N=10) examined design opportunities and tensions.The review considered how narrative framing, literary interventions, and imperfect LLM outputs could support or disrupt reflection, voice, and meaning-making.Source: §1 · p. 3

2 Related Work

The related work connects personal storytelling to meaning-making and traces restorying across literary practice and psychology. It positions HWAMW against narrow, static interventions and AI writing approaches that may homogenize narratives or steer personal writing toward particular mental states.

2.1 Restorying

  • People use narratives to organize experience and explore meaning, making the stories they tell relevant to their sense of identity and life meaning.Source: §2.1 · p. 3

2.1.1 In literature and creative writing.

  • Literary restorying ranges from adaptation to reinterpretation, and metafiction offers a frame for exploring relationships among stories and broader meanings.Source: §2.1.1 · p. 3

2.1.2 Psychological effects of restorying.

  • Psychological restorying changes how people tell events, not the events themselves; prior work found a hero’s-journey intervention could increase perceived meaning in life.Source: §2.1.2 · p. 4
  • Prior psychological restorying research centers on the hero’s journey and static prompts, motivating broader storytelling techniques and more dynamic interventions.Source: §2.1.2 · p. 4

2.2 Large Language Models and Narrative Writing

  • LLM narratives can appear convincing yet differ from human writing in style and narrative choices, amid concerns about replacing human writing and homogenizing creative outputs.Source: §2.2 · p. 4

2.2.1 AI Writing Assistants.

  • AI writing assistants may support accessibility, productivity, and creative expression, while writers’ ethical concerns and values—including creativity and authenticity—shape their reception.Source: §2.2.1 · p. 4

2.2.2 LLM Interfaces for Personal Writing.

  • Unlike reflective systems that guide experiences toward positive themes, literary restorying can dwell on complications; HWAMW treats personal narratives as valuable beyond their effects on mental states.Source: §2.2.2 · p. 4

3 The House with a Million Windows

HWAMW places personal-story restorying in an interactive metafictional house, where users explore LLM-generated literary variations and then retell their story in their own words. Its design emphasizes flexible interventions, user agency, and transparent presentation of the literary influences behind generated variations.

3 The House with a Million Windows

  • HWAMW guides users to tell a personal story, explore fictionalized LLM variations, reflect, and write the story again in their own words.Source: §3 · p. 4

3.1 Restorying Interventions, Redesigned

  • The redesign combines metafictional framing, flexible literary interventions, dynamic retellings, in-story and questionnaire measures of meaning, and user choice over which changes to adopt.Source: §3.1 · p. 5

3.2 Reframing AI assistance in personal writing

  • Because LLMs can generate story variations rapidly and flexibly, the system frames them to preserve users’ authorship and authenticity throughout restorying.Source: §3.2 · p. 5

3.2.1 Metafictional Framing.

  • The house’s container story lets users coauthor the frame and their personal narratives, supporting agency within a potentially vulnerable writing process.Source: §3.2.1 · p. 5

3.2.2 AI outputs as “windows” onto existing narratives.

  • The “window” metaphor presents LLM variations as non-authoritative views onto a user’s story, foregrounding literary influence and avoiding chatbot anthropomorphism.Source: §3.2.2 · pp. 5–6

3.3 System Architecture

  • The web prototype uses Llama-3.3-70B-Instruct and stores interaction data, while its current version permits one play-through per participant and no return to earlier steps.Source: §3.3 · p. 6

3.3.1 Windows.

  • Each window prompt pairs a public-domain literary work with style and structure/content rules; the prototype uses five author-developed examples, with plans to diversify and ground them in literary expertise.Source: §3.3.1 · p. 6

3.4 User Experience

  • Users tell a story and rate its meaningfulness, explore three selected windows with opportunities to request changes and view inspirations, then retell their story and rate it again.Source: §3.4 · p. 6, Figure 3 · p. 7

4 Empirical Validation

In a within-subjects study, participants’ self-reported narrative identity increased after using HWAMW, alongside changes in their retold stories. Participants also rated the experience positively, while the story analyses found higher staging and plot progression but no increase in cognitive tension.

4.1 Recruitment

  • The within-subjects study recruited 40 UK or Ireland residents with at least a high school diploma, without screening for narrative or AI expertise.Participants were recruited through Prolific.Source: §4.1 · p. 7, §4 · p. 7

4.2 Self-report measures and procedure

  • The study compared participants’ NISE responses and writing before and after HWAMW, using LIWC-22 to analyze the stories.NISE was the sole survey instrument; the researchers chose it for its thematic relevance.Source: §4.3 · p. 7, §4.2 · p. 7

4.3 Results

  • NISE scores rose significantly from 4.55 before HWAMW to 4.75 after it (t(39) = 4.24, p < 0.001, Holm-corrected, dz = 0.67).The increase appeared driven primarily by autobiographical reasoning and positive themes.Source: §4.3 · pp. 7–8, Figure 4 · p. 8
  • Participants rated HWAMW as helpful for thinking about personal stories (87.3/100), enjoyable (83.8), thought-provoking (78.6), and worth using again (76.3).Thirty-seven of 40 participants could name someone or a group with whom they might share it.Source: §4.3 · p. 8
  • Time 2 stories scored higher than Time 1 on staging (21.27 to 32.62) and plot progression (9.16 to 28.27), while cognitive tension did not increase.Eight story pairs below the 100-word minimum were excluded, leaving 32 pairs.Source: §4.3 · p. 8, Table 2 · p. 9
  • Across all 40 participants, VADER found more positive Time 2 narratives, while SentiArt showed increases across all six emotional dimensions.The SentiArt pattern suggests more affective content overall, not only more positive content.Source: §4.3 · p. 8, Table 1 · p. 8

5 Expert Review

An expert review examined how users experienced HWAMW and how its design might be improved. Experts described expanded perspectives and retained ownership during restorying, valued the narrative framing, and identified both possibilities and limitations in the generated windows.

5.2 Procedure

  • The review recruited 10 postgraduate-level experts in AI and narrative or adjacent fields and analyzed structured free responses using template analysis.Experts did not complete pre-interface self-report measures.Source: §5.1 · p. 9, §5.2 · p. 9

5.3.1 Restorying with HWAMW challenged and expanded users’

  • Experts said iterative windows helped them work through writing blocks, consider more than their initial telling, and see their stories from new perspectives while retaining ownership.Some reported that their story’s meaning shifted, yet the final story still felt like their own.Source: §5.3.1 · p. 9

5.3.2 HWAMW’s narrative framing made it easier to engage.

  • Experts found the metafictional setup engaging and said its narrative framing made sharing personal or vulnerable stories feel more natural.They described it as guided engagement without therapy’s explicit guidance.Source: §5.3.2 · p. 9

5.3 Results

  • Experts enjoyed windows inspired by iconic fiction, but recommended broader genres, forms, and writers, including works that challenge a limiting literary canon.They also proposed more radical changes to emotional tone, temporality, plot, and setting.Source: §5.3.3 · p. 9

5.3.4 LLM limitations were obvious to every user, but surprisingly

  • Some experts found generated windows predictable, clichéd, or overwrought, while others found the excess entertaining.For some, the style was obtrusive, annoying, or offensive given their story.Source: §5.3.4 · pp. 9–10
  • Experts differed on counterfactuals and hallucinations: some preferred faithful retellings, while others found mismatches useful for reflection.One expert said tonal mismatches helped locate their emotional position more precisely.Source: §5.3.4 · p. 10

6 Discussion

The discussion positions HWAMW as a way to address concerns that AI-assisted writing can diminish human storytelling. It argues that literary restorying offers a broader set of possible tellings, while framing LLMs as a means of encountering and reconsidering human narratives rather than replacing them.

6 Discussion

  • The authors present HWAMW as one possible design response to concerns that AI-assisted writing can strip meaning and intentionality from storytelling.Their empirical findings suggest users re-encountered and reinterpreted stories, though the discussion does not claim a direct comparison with prior psychological findings.Source: §6 · p. 10
  • Unlike a single intervention such as the Hero’s Journey, HWAMW offers varied literary styles and techniques, potentially connecting users to new perspectives and cultural meanings.The authors note that a single structure may not suit every story.Source: §6 · p. 10
  • The authors frame LLMs as counternarratives for reconsidering human stories, potentially giving people more agency over stories told about them and about AI.They propose using generated text to make structures, assumptions, and possibilities in human stories newly visible.Source: §6 · p. 10

7 Limitations and Future Work

The prototype’s findings and scope are bounded by its limited set of literary windows, narrow psychological measurement, text-only design, and focus on intrapersonal meaning. Future work could broaden the interventions, evaluations, modalities, and target outcomes.

7 Limitations and Future Work

  • The prototype includes few windows drawn from a limited literary canon, and prioritizes an integrated experience over precise measurement of individual windows’ effects.Future windows could broaden authors, genres, forms, and narrative techniques, with domain experts co-designing presets and users designing and testing their own.Source: §7 · p. 10
  • Psychological validity is assessed with one self-report scale, and further research is needed to interpret the findings and determine their practical meaning.Proposed work includes other scales, window-by-window comparisons, richer qualitative questions, closer readings of restoried texts, and study of effects’ durability and possible links to design, framing, format, and windows.Source: §7 · pp. 10–11
  • The current prototype is text-only; future iterations could explore immersive visual design, spatial navigation, and spoken-story inputs and outputs.These changes could make the system more accessible to people who find longform writing difficult or intimidating, and may change the stories they tell and their restorying experience.Source: §7 · p. 11
  • The prototype focuses on intrapersonal meaning, while possible future applications include cultural meaning and stories’ real-world impacts.Examples include perspectives shaped by different backgrounds or worldviews, and restorying narratives about climate change, technology, or misinformation.Source: §7 · p. 11

Ethics and Privacy Statement

HWAMW is a narrative experience, not therapy, and has no clinical aims or oversight. Its unpredictable LLM outputs and risk of misuse warrant safeguards and warnings in deployments.

Ethics and Privacy Statement

  • HWAMW is not therapy and has no clinical aims or oversight; its unpredictable LLM outputs risk misinterpretation and misuse.The authors say it should not be used by psychologically vulnerable populations or treated as a therapy tool, and ask deployers to warn users and safeguard against misuse.Source: Ethics and Privacy Statement · p. 11

A Example User Walk-through

The section introduces a walkthrough of the narrative using pseudonymised, occasionally abbreviated real inputs from one user, called Ahmed.

A Example User Walk-through

A.1 The Approach

Ahmed’s journey moves from reflective prompts in the woods to telling a true story at the House, then reviewing its summary and considering its meaning.

A.1 The Approach

  • In the woods, prompts ask Ahmed to list his worries and describe himself, providing context while also prompting reflection on the story he wants to tell.He selects concerns including his career, creative pursuits, and community, and describes himself as caring, affectionate, and cheerful.Source: §A.1 · p. 12
  • At the House, the Gatekeeper asks Ahmed for a true story with a beginning, middle, and end, capped at 300 words.Ahmed recounts his path from football interest to managing a team, stepping aside for his assistant, and seeing the team win.Source: §A.1 · p. 13
  • After Ahmed submits the story, the Gatekeeper presents an LLM-generated plot summary for confirmation, asks for character details, and asks about his motivations.The system prompts for details about up to three important characters besides the user; Ahmed adds information about his assistant manager and selects motivations including supporting his family, friends, and community.Source: §A.1 · p. 13
  • The Gatekeeper asks Ahmed to rate his story’s meaning on a seven-notch scale, then asks his name and preferred pronouns and welcomes him into the House.The scale runs from “A story that changed everything” to “A story still waiting to make sense”; Ahmed selects the third notch, just above the center.Source: §A.1 · p. 13

A.2 The House

In the House, Ahmed encounters windows that turn his story into literary-style variations. He can explore several windows before dawn ends the visit.

A.2 The House

  • The House is a gallery-like space of distinctive, oddly lit windows, which Ahmed approaches after wandering inside.Source: §A.2 · p. 13
  • The first window presents a Pride and Prejudice-inspired football story in three 100–150-word sections, then reveals its source and style rules.The rules specify witty, elevated Regency diction, third-person free indirect discourse, and verbal irony.Source: §A.2 · p. 13
  • The Pride and Prejudice-style plot begins with disadvantage, worsens before resolution, and uses dialogue, changing social situations, and mistaken judgments.Questioning assumptions brings growth and resolution.Source: §A.2 · p. 13
  • Ahmed explores two more windows after the first, choosing each from three options he has not yet seen.The sequence ends when dawn arrives and he steps outside.Source: §A.2 · p. 13

A.3 The Departure

At departure, Ahmed retells his story in his own words after reviewing the windows. The following examples specify literary styles and structures used for the windows.

A.3 The Departure

  • Ahmed’s final retelling is longer and more descriptive, develops a clearer conflict and ending, and he places it at the top of the scale.He can revisit the window texts, but the prompt emphasizes that the new story should be his own.Source: §A.3 · pp. 13–14

B.1 “Hills Like White Elephants” by Ernest Hemingway

  • The “Hills Like White Elephants” window uses plain, short dialogue to imply tension and an indirectly resolved decision tied to a physical journey.Source: §B.1 · p. 14

B.2 To the Lighthouse by Virginia Woolf

  • The To the Lighthouse window combines cumulative, stream-of-consciousness prose and recurring imagery with natural details and shifting character perspectives.Source: §B.2 · p. 14

B.3 “The Tell-Tale Heart” by Edgar Allen Poe

  • The Tell-Tale Heart window uses a retrospective, increasingly unstable first-person narrator, whose fixation drives tension toward a decisive climax.Its rules call for formal, archaic language and dark Gothic imagery.Source: §B.3 · p. 14

B.4 Pride and Prejudice by Jane Austen

  • The Pride and Prejudice window uses witty Regency diction, free indirect discourse, and irony alongside social misunderstandings that prompt reconsideration.Source: §B.4 · p. 14

B.5 “Salvation” by Langston Hughes

The “Salvation” window combines a conversational, retrospective voice with social commentary and multiple character perspectives. A future iteration will use only public-domain works accessible to participants in deployment countries.

B.5 “Salvation” by Langston Hughes

  • Future iterations will use works that are public domain and accessible to participants in deployment countries.Source: §B.5 · p. 14
  • The “Salvation” style rules call for conversational irony, belief-and-disillusionment imagery, distinct minor-character voices, social commentary, and retrospective narration.The narrator reflects on events in past tense while showing other characters’ internal conflicts.Source: §B.5 · p. 14
Loading 2609.12537v1…