Figure 1: Our real-time interactive storytelling system in a live theater environment. The system dynamically aggregates raw audience chat inputs to shape an ongoing audiovisual narrative, using AI-generated performances optimized to maximize audience laughter.
Abstract
Large language models can generate joke-like text, but making a live audience laugh within an audiovisual story requires more than isolated humor generation.
We present a live crowd-to-story system that automatically groups and prioritizes pseudonymous audience comments and integrates selected threads into approximately one-minute audiovisual story blocks. The system clusters related comments, prioritizes them by narrative fit and audience response, preserves recognizable audience phrasing through near-verbatim “whispers,” and stores successful ideas in memory so they can return as recurring jokes and callbacks. This design creates a rapid feedback loop in which participants can see their contributions acknowledged by both the crowd and the story.
Across 133 stories, 401 participants, and 793 person-in-story observations, early inclusion of a participant’s comment in played story content was associated with a 51.7% adjusted difference in the geometric mean of subsequent commenting. Comment-linked dialogue and on-screen whispers showed positive adjusted differences of 38.1% and 28.1%, respectively. These observational results do not establish causality, but they support the design principle that rapid, recognizable inclusion can reinforce participation. We argue that comic payoff in live generative storytelling emerges through distributed agency among the audience, AI writers, platform, and human creators.
133 stories401 participants793 person-in-story observations+51.7% adjusted difference with early inclusion
Section 1
Introduction
Generating joke-like text is different from making a live audience laugh. In an ongoing audiovisual story, humor depends not only on the words being spoken, but also on narrative context, character behavior, delivery, timing, and the audience’s shared understanding of what has happened before. This challenge becomes more complex when the audience is allowed to influence the story in real time: the system must respond quickly enough for participants to recognize their influence, while remaining selective enough to preserve a coherent narrative.
Humor depends not only on
the words being spokenbut also onnarrative contextcharacter behaviordeliverytimingthe audience’s shared understanding of what has happened before
At the same time, a live audience offers something that generative models lack: a continuous source of unexpected, socially grounded comic material. Audience members can propose unusual ideas, react to one another, and develop successful premises through riffs and variations. This motivates a different division of creative labor from conventional AI humor generation. Rather than asking the AI to independently invent the jokes, we investigate a setting in which humans supply much of the comic material while AI writers select, structure, remember, and perform it as part of a story.
Humans supply
much of the comic material
AI writers
selectstructurerememberperform
it as part of a story
We study this setting through a live crowd-to-story system in which participants submit pseudonymous comments while watching an evolving audiovisual narrative. The system groups related contributions into threads, selects a small number for each approximately one-minute story block, preserves recognizable audience phrasing through near-verbatim whispers, and stores successful ideas so they can return as recurring premises or callbacks. This work addresses three main questions:
1
How can contributions from a live crowd be incorporated rapidly into an audiovisual story without sacrificing narrative and character coherence?
To investigate these questions, we analyze Creator Test and Production sessions. This paper presents a live crowd-to-story architecture that transforms audience input into short audiovisual narrative blocks, together with an observational analysis of how recognizable story inclusion relates to subsequent participation. We also examine the experience through the lens of distributed agency, considering how creative control is shared among the crowd, the AI system, human creators, and the platform.
Section 2
Background
Interactive and co-creative storytelling.
Interactive narrative systems seek to balance participant agency with authorial control. Façade [1] selects from pre-authored story beats in response to player actions, whereas IDtension [2] generates narrative events dynamically from an explicit model of characters, goals, and story constraints. These systems made participation consequential, but their narrative possibilities remained largely bounded by pre-authored actions, dialogue, or branches. Recent LLM-based systems have expanded this design space by enabling dialogue and narrative events to be generated dynamically. Dramatron [3] generates long-form scripts through hierarchical generation of characters, plot beats, scenes, and dialogue , while generative-agent architectures use memory, reflection, and planning to maintain behavior over extended interactions [4, 5].
Façade [1]selects from pre-authored story beats
IDtension [2]generates events from an explicit model of characters, goals, and constraints
Dramatron [3]hierarchical generation of characters, plot beats, scenes, and dialogue
Generative agents [4, 5]memory, reflection, and planning over extended interactions
This worka live crowd, multiple AI writers, and a predefined story framework
pre-authoredgenerated dynamically
Collective creative systems distribute creative decisions between humans and computational agents rather than assigning control exclusively to either party [6]. In such systems, AI may propose, interpret, extend, or transform human input, while human participants contribute intent, judgment, and domain knowledge [7, 6]. Our setting extends this idea from collaboration between an individual creator and an AI system to collaboration among a live crowd, multiple AI writers, and a predefined story framework. Audience members continuously propose material while watching the same evolving story, and the system must select and return their contributions quickly enough for that influence to remain recognizable.
Prior setting
an individual creator and an AI system
Our setting
predefined story framework
a live crowd, multiple AI writers, and a predefined story framework
Computational humor and audience response.
Computational humor research has treated humor as a property of an individual text, focusing on joke and pun generation, humor detection and classification, punchline generation, or offline ratings of whether one sentence is funnier than another [8, 9, 10]. Other work has incorporated audience characteristics, showing that humor preferences can vary across demographic groups [11]. These approaches provide useful measures of linguistic humor, but they do not capture the broader conditions under which a performed narrative produces laughter.
In a live audiovisual story, comic payoff may depend on prior events, character consistency, delivery timing, audience reactions, and the repeated development of a shared premise. The same line may therefore succeed in context and fail in isolation. Laughter is also social: participants observe others’ responses, build on recognizable jokes, and may find humor in the audience chat before an idea enters the story. The relevant object of analysis is thus not an isolated text, but a temporally extended interaction among narrative context, performance, and collective audience behavior.
Recent Creative AI work has explored computational spectators that simulate internal dialogue and affective responses to media [12]. Our work instead centers a real human audience that both produces and responds to the comic material. We study how a human–AI system transforms crowd-originated ideas into coherent narrative events: the crowd proposes and socially develops the material, while the AI system selects, integrates, performs, and remembers it. Our central hypothesis is that humor emerges from this interaction, particularly when audience contributions return quickly, remain recognizable, and develop into recurring threads or callbacks.
Section 3
LaughLoop: The Crowd-to-Story System
Our system transforms live audience contributions into a coherent audiovisual story. Audience comments are continuously grouped and prioritized by an automated AI pipeline, then selected threads are incorporated into successive story blocks of approximately one minute. This short generation cycle allows audience input to affect the performance while remaining recognizable to participants, as illustrated in Fig. 2.
Figure 2: Crowd-sourced laughter from LaughLoop. A closed ~1-minute content loop turns audience ideas into story canon, social proof and callbacks compound laugh.
Live audience environment.
A browser interface presents the generated story alongside a shared audience chat, allowing viewers to follow the performance, propose ideas, and react to emerging joke threads. The chat therefore functions as both an input channel and part of the entertainment experience. Participants use persistent pseudonymous aliases. This reduces real-world social exposure and can encourage more playful contributions, while preserving attribution through reactions, replies, and recognizable story inclusion. Early seeded jokes and playful comments establish the expected tone and encourage participants to contribute unconventional ideas.
browser interface
generated story
shared audience chat
persistent aliaspropose ideas
persistent aliasreact to emerging joke threads+3
seededearly seeded jokes establish the tone
Schematic. The chat is both an input channel and part of the entertainment experience.
Audience comment processing.
Audience comments arrive continuously and at a higher rate than can be directly incorporated into the story. Therefore, a low-latency LLM with limited narrative context filters irrelevant, unsafe, repetitive, or scene-incompatible suggestions, clusters related comments into candidate threads, and prioritizes them by narrative fit, audience support, and potential for continued interaction. Reactions, replies, and repeated variations provide social evidence that a thread is gaining recognition. The system does not attempt to include every suggestion. Doing so would fragment the narrative, weaken character consistency, and dilute the strongest comic ideas. Instead, it selects a number of promising threads that can be developed without pulling the story in unrelated directions.
Comments arrivecontinuously, faster than they can be incorporated
Prioritizenarrative fit · audience support · potential for continued interaction
Selecta number of promising threads
A low-latency LLM with limited narrative context performs the filtering, clustering, and prioritizing.
Story-block generation.
For each approximately one-minute block, the automated writing system receives the current narrative state, character information, recent dialogue, prioritized audience threads, and creator-authored story DNA. The story DNA defines the overall narrative arc, character backstories, world rules, and lightweight branch dependencies. Each block generally draws from no more than two audience threads, allowing them to influence dialogue, character actions, and plot development without losing narrative focus. At least one selected contribution is incorporated near-verbatim as a whisper, keeping its source recognizable to the contributor and wider audience. The system then generates dialogue and story beats together with voices, facial-animation parameters, character actions, and game-engine commands. These commands are rendered as live video and returned to the audience, keeping the interval between a contribution and its narrative payoff short.
Receives
current narrative state
character information
recent dialogue
prioritized audience threads ≤ 2 per block
creator-authored story DNA
narrative arc · character backstories · world rules · lightweight branch dependencies
Automated writing system
one block ≈ 1 minute
Generates
dialogue and story beats
≥ 1 near-verbatim whisper
voices
facial-animation parameters
character actions
game-engine commands → live video
Narrative and comedic memory.
The system maintains a canonical narrative state alongside a memory of audience-derived threads. The canonical state tracks persistent changes to characters, relationships, objects, and plot events, while the audience memory records which ideas have been selected, how they appeared, and how they affected the story. Once incorporated, an audience contribution can become part of the story canon and later return through dialogue, character behavior, or visual details. This allows an isolated suggestion to develop into a recurring premise or callback, while tracking which audience ideas have already influenced the narrative.
Canonical narrative state
tracks persistent changes to
charactersrelationshipsobjectsplot events
Audience memory
records
which ideas have been selectedhow they appearedhow they affected the story
isolated suggestion→story canon→returns through dialogue, character behavior, or visual details→recurring premise or callback
Feedback and audience credit.
Audience influence is made visible through three forms of feedback: immediate reactions, replies, or extensions from other peer viewers; recognizable inclusion in dialogue, an on-screen whisper, or a character action; and delayed recognition when an idea returns as a callback. Persistent aliases and recognizable reuse allow participants to trace a contribution from the chat into the performed story. Figure 3 illustrates this progression from peer feedback to narrative recognition and continued collective development. Section 4 examines whether early story inclusion is associated with later participation.
Audience feedback as a reinforcing narrative loop
Figure 3: Conceptual audience-engagement trajectory following an audience contribution. Peer reactions provide immediate social validation, recognizable inclusion in the generated story provides narrative recognition, and continued audience riffing can establish a reinforcing callback loop.
AI resistance and comic tension.
Audience contributions are treated as narrative proposals rather than direct commands. Characters may resist, misunderstand, delay, or only partially accept a popular suggestion, since immediate compliance can feel mechanical, weaken character consistency, and remove comic tension. The writing system therefore balances responsiveness with the character’s established goals and personality, allowing audience influence to remain recognizable without removing the fictional world’s autonomy.
A popular suggestion is a narrative proposal, not a direct command
responsivenessthe character’s established goals and personality
Immediate compliance can feel mechanical, weaken character consistency, and remove comic tension.
Section 4
Live Audience Study
We analyzed whether recognizable audience influence on the generated story was associated with continued participation, measured by subsequent commenting.
Data and Cohort.
The analysis includes Creator Test and Production sessions conducted between May 1 and July 29, 2026. We consider multi-person stories with at least two participants identified through entry records or audience comments, yielding 133 stories, 401 unique participants, and 793 person-in-story observations. Each observation represents a participant who commented during the first half of a story. We define early story inclusion as a participant-linked comment appearing in played content between 25% and 50% of story progress, and measure subsequent participation by comments submitted by the same participant between 50% and 100%. Analyses of audience size use the 57 stories with complete entry records.
133stories
401unique participants
793person-in-story observations
57stories with complete entry records
Story progress · May 1 – July 29, 2026
0–25%25–50%early story inclusion50–100%subsequent participation
0%25%50%100%
Each observation is a participant who commented during the first half of a story.
Tracing Audience Influence.
Audience contributions may influence generation through grouped threads, sentiment summaries, saved story context, or retrieved memory, so not every contribution can be linked to a specific played line. Across 95 stories, the system recorded 6,543 participant-linked thread contributions, of which 6,522 (99.7%) entered saved story context and 462 (7.1%) were linked to played dialogue or an on-screen whisper. Comparing 44,668 story-writing inputs with final dialogue and whispers identified 1,007 strict textual matches. This is a conservative measure of direct reuse because paraphrased contributions or ideas affecting plot, tone, or character behavior may not produce exact matches. The system also recorded 5,751 story-memory references, 1,458 memory lookups, and 6,535 saved-context entries. Because individual memory retrievals cannot be linked reliably to specific played lines, these counts indicate that memory was available during generation rather than how often it directly affected the final story.
Linked to played dialogue or an on-screen whisper462 7.1%
1,007strict textual matches among 44,668 story-writing inputs
5,751story-memory references
1,458memory lookups
6,535saved-context entries
Outcome and Statistical Model.
The primary outcome is later commenting activity, yi,s = log(1 + Clatei,s), where Clatei,s is the number of comments submitted by participant i after 50% story progress in story s. We estimate
Z audience-influence variableCearly early comment volumeTfirst first-comment timingRearly early reactions receivedαs story fixed effects
where Zi,s is the audience-influence variable, Cearlyi,s is early comment volume, Tfirsti,s is first-comment timing, Rearlyi,s is the number of early reactions received, and αs denotes story fixed effects. We evaluate four influence variables separately: any comment-linked story content, linked dialogue, linked on-screen whispers, and reactions to an early comment. Estimates are reported as adjusted percent differences, 100(exp(β1) − 1), with confidence intervals clustered by story.
Table 1: Association between early audience recognition and later commenting. Estimates are adjusted percent differences in the geometric mean of 1+ second-half comments.
Early recognition signalParticipantsAdjusted differencep-value
Any comment-linked story content164+51.7%< .001
Comment-linked dialogue153+38.1%.005
Comment-linked on-screen whisper89+28.1%.009
Each reaction to an early comment43+21.8%.094
Played story inclusion · confidence intervals above zero Peer reactions alone · confidence interval crossed zero
Early Inclusion and Later Participation.
Table 1 reports the associations between early audience recognition and subsequent commenting. All three forms of played story inclusion were positively associated with later commenting, with confidence intervals above zero. Reactions to early comments also had a positive estimate, but its confidence interval crossed zero, providing stronger evidence for recognizable story inclusion than for peer reactions alone. Because the outcome is log-transformed, these estimates describe adjusted differences in the geometric mean of 1+ later comments, not equivalent increases in raw average comments for every participant.
Audience Scale and Participation Dynamics.
Across the 57 stories with complete audience-size records, audience entrants strongly predicted total comment volume when sessions were pooled (R2 = .909), particularly in Production (R2 = .973), but not in Creator Test sessions (R2 = .002). Comments per entrant did not increase monotonically with audience size (ρ = .103, p = .446), indicating that larger audiences generated more comments overall without necessarily increasing individual activity. Participation declined over time: the ratio of distinct comment-author aliases to the measured active audience fell from 54.6% in the first half to 23.5% in the second, a difference of −31.1 percentage points with a 95% interval of [−39.5, −23.9]. Because some Production aliases cannot be matched reliably to entrant identities, this is a count ratio rather than the percentage of known individuals who commented.
Entrants → total comment volume · R²
.909Pooled
.973Production
.002Creator Test
57 stories with complete audience-size records. Comments per entrant: ρ = .103, p = .446.
Distinct comment-author aliases ÷ measured active audience
−31.1 percentage points, 95% interval [−39.5, −23.9]. A count ratio, not the percentage of known individuals who commented.
Interpretation.
These results are observational. More engaged participants may contribute ideas that are selected and remain active throughout the story. The study measures later commenting rather than laughter, or perceived authorship. Direct evaluation of audience laughter remains future work.
Section 5
Discussion
Our findings suggest that laughter in a live generative story emerges from a social and narrative process rather than from isolated joke generation. Audience members propose and develop ideas, while the AI writing system selects, structures, and performs them within a coherent fictional world.
5.1 What Makes the Audience Laugh?
Social permission and recognition.
Audience members may hesitate to contribute unusual or risky ideas until the environment signals that playful participation is acceptable. Seeded jokes and visible examples of successful contributions help establish this tone, while pseudonymous aliases reduce real-world exposure without removing attribution. Recognition then provides evidence that participation matters. Participants whose comments appeared in played story content tended to remain more active later. Peer reactions offer immediate social validation, but recognizable inclusion in dialogue, a character action, or an on-screen whisper provides a more consequential form of acknowledgment.
Riffing is easier than invention.
Audience humor rarely emerges from isolated jokes; participants build on ideas that have already gained attention. A recognizable premise gives others shared material to modify, escalate, contradict, or reframe, reducing the need to invent a new joke. The system supports this by clustering related comments into threads and preserving successful ones in narrative memory. Once incorporated into the story, threads can return as callbacks, enabling continued development while maintaining continuity across story blocks.
Repetition creates shared comic history.
Repetition is most effective when each return develops the premise rather than merely duplicating it. A joke may begin in the chat, enter dialogue, shape a character’s beliefs, and later return as a callback, showing that the fictional world has absorbed it. This progression also distributes authorship: one participant introduces the premise, others vary it, and the AI writing system turns those variations into a recurring narrative thread. The resulting laugh therefore grows from a shared history rather than a single generated line.
chat→dialogue→character’s beliefs→callback
Selectivity protects both humor and coherence.
The system cannot incorporate every audience suggestion, especially in large crowds where ideas may conflict or poorly fit the scene. Indiscriminate inclusion would weaken coherence and make individual contributions less recognizable. Filtering and prioritization therefore serve both creative and technical roles by selecting a small set of threads that can develop coherently over time. This also creates an authorship asymmetry: the audience contributes freely, but the platform determines how comments are grouped, which signals indicate quality, and which ideas enter the story.
Character resistance can strengthen comic payoff.
Immediate compliance with an audience suggestion is not always the funniest response. When characters simply execute popular commands, the interaction can feel mechanical and the fictional world may lose its internal logic. Partial resistance, misunderstanding, or delay can preserve character consistency while creating anticipation. The AI writing system therefore reflects the audience’s comic intention without making characters fully obedient. A character may reject, misunderstand, or reluctantly engage with a premise before eventually accepting it, making the resulting payoff feel more earned and consequential.
Crowd size supports opportunity, but not automatically quality.
Practitioner observations suggest that the format becomes more reliable with roughly 25–30 participants, as larger crowds create more opportunities for unusual ideas, reactions, and collaborative riffing, although our analysis does not identify a specific threshold. In Production sessions, audience size was strongly associated with total comment volume, but comments per entrant did not increase monotonically. Larger crowds therefore expand the pool of material without necessarily increasing individual activity or humor quality; crowd composition, social tone, and a few highly generative participants may matter as much as size.
Agency: Who Gets Credit for the Laugh?
A successful comic moment cannot be attributed to a single creator because creative agency is distributed across the audience, AI system, human creators, and platform.
The audience contributes premises, phrases, and unexpected associations, evaluates them through reactions, and extends them through replies and variations.
The AI writing system selects and reframes these contributions, determines how they enter the scene, assigns them to characters, and decides whether they return later.
Human creators define the story world, characters, narrative arc, safety boundaries, and tone, shaping the space in which audience and model agency operate.
The platform influences authorship by clustering comments, counting reactions, filtering threads, managing memory, and determining how quickly audience ideas affect the story.
This distribution is not equal. Highly active participants may dominate emerging threads, majority reactions may suppress less common forms of humor, and automated ranking may favour ideas that are easier to classify or integrate. The system may therefore both reflect audience taste and gradually steer participation toward the forms of humor it can process most effectively. We consequently describe the resulting humor as crowd-sourced rather than democratically authored. The audience supplies much of the comic material, but that material is filtered and transformed by the AI system, bounded by creator-defined rules, and mediated by the platform. The resulting laughter emerges from the interaction among these agents rather than belonging to any one of them.
mediated by the platform
bounded by creator-defined rules
filtered and transformed by the AI system
the audience supplies much of the comic material
Crowd-sourced rather than democratically authored.
Section 6
Conclusion
Making a live audience laugh at an AI-generated story requires more than isolated jokes. Our system creates a rapid loop in which audience members contribute comic ideas, AI writers integrate selected threads into a coherent audiovisual narrative, and recognizable inclusion encourages continued participation and collective riffing. Observational results suggest that participants whose comments appeared in story content were more likely to remain active later, though this association should not be interpreted causally. More broadly, humor in live generative storytelling emerges through shared agency among the audience, AI writers, platform, and human creators. The central design challenge is not simply to maximize laughter, but to preserve narrative coherence while making audience influence visible, timely, and meaningful.
References
Michael Mateas and Andrew Stern. Structuring content in the Façade interactive drama architecture. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 1(1):93–98, 2005.
Nicolas Szilas. IDtension – highly interactive drama. In Proceedings of the Fourth Artificial Intelligence and Interactive Digital Entertainment Conference, volume 4, pages 224–225, 2008.
Piotr Mirowski, Kory W. Mathewson, Jaylen Pittman, and Richard Evans. Co-writing screenplays and theatre scripts with language models: Evaluation by industry professionals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, New York, NY, USA, 2023. Association for Computing Machinery.
Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, New York, NY, USA, 2023. Association for Computing Machinery.
Parag K. Mital, Seth Rosetter, Arturo Castro Prieto, Jacobo Heredia Zurita, and Breanna Browning. Orchestrating emergent storytelling with embodied multi-agent systems. In Proceedings of the 39th Conference on Neural Information Processing Systems, Creative AI Track, 2025.
Zhiyu Lin, Upol Ehsan, Rohan Agarwal, Samihan Dani, Vidushi Vashishth, and Mark O. Riedl. Beyond prompts: Exploring the design space of mixed-initiative co-creativity systems. In Proceedings of the 14th International Conference on Computational Creativity, pages 64–73, 2023.
Georgios N. Yannakakis, Antonios Liapis, and Constantine Alexopoulos. Mixed-initiative co-creativity. In Proceedings of the 9th International Conference on the Foundations of Digital Games. Society for the Advancement of the Science of Digital Games, 2014.
Miriam Amin and Manuel Burghardt. A survey on approaches to computational humor generation. In Proceedings of the 4th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 29–41, Online, 2020. International Committee on Computational Linguistics.
Nabil Hossain, John Krumm, Michael Gamon, and Henry Kautz. SemEval-2020 task 7: Assessing humor in edited news headlines. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 746–758, Barcelona, Spain (Online), 2020. International Committee for Computational Linguistics.
Orion Weller and Kevin Seppi. Humor detection: A transformer gets the last laugh. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3621–3625, Hong Kong, China, November 2019. Association for Computational Linguistics.
Aparna Garimella, Carmen Banea, Nabil Hossain, and Rada Mihalcea. “judge me by my size (noun), do you?” YodaLib: A demographic-aware humor generation framework. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2814–2825, Barcelona, Spain (Online), 2020. International Committee on Computational Linguistics.
Pat Pataranutaporn, Chayapatr Archiwaranguprok, Manaswi Mishra, and Pattie Maes. Artificial spectators: Developing AI-generated audiences for AI-generated content—a speculative exploration of AI, media, and the attention economy. In Proceedings of the 39th Conference on Neural Information Processing Systems, Creative AI Track, 2025.
Citation
BibTeX
@inproceedings{adeli2026laughloop,
title = {How to Get an Audience to Laugh at a Live Generative Story:
Crowd-Sourced Humor and Shared Agency},
author = {Adeli, Vida and Schreck, Benjamin and Mehraban, Soroush and Clifford, Cole},
booktitle = {NeurIPS 2026 Creative AI Track},
year = {2026},
url = {https://openreview.net/forum?id=iAb96aSsAn}
}