PaperlayerView exact revision on arXiv

Source-linked AI summary

Collaborative Storytelling with Human Actors and AI Narrators

Boyd Branch, Piotr Mirowski, Kory W. Mathewson

arXiv:2109.14728v1cs.HCcs.AI

TL;DR

The paper asks how AI language models can collaborate with human performers in long-form improvised storytelling while tracking plot progression and character arcs. It stages GPT-3 as a live narrator working with professional improvisers and evaluates performances through transcripts and surveys. Audiences and performers responded positively, while the study emphasizes human curation and improvisational theatre as the collaboration’s practical boundary and test-bed.

  • Problem

    AI systems must track plot progression and character arcs when collaborating with human actors, but prior collaborative storytelling work often centered on virtual environments or AI characters.

    Source: Improv Theatre and AI · p. 1

  • Method

    The study used GPT-3 as a live narrator with automated content filters, human operator selection, an avatar, two public performances, transcript analysis, and surveys of performers and audience members.

    Source: Avatar for the AI narrator · p. 2, Evaluating AI in performance · p. 3, Results · p. 3

  • Results

    AI narration generated novel plot directions that advanced story arcs and received positive responses from audiences and performers.

    Source: Discussion · p. 4, Conclusions · p. 4

  • Takeaways & Limitations

    Improvisational theatre is a useful test-bed for exploring human-AI creative collaboration, with human interpretation and curation remaining part of the storytelling process.

    Source: Discussion · p. 4, Conclusions · p. 4

    Because language models reproduce patterns, biases, and misalignment from their training data, the system required automated filtering and human curation of generated content.

    Source: Avatar for the AI narrator · p. 2, Conclusions · p. 4

Abstract

from arXiv · show

Large language models can be used for collaborative storytelling. In this work we report on using GPT-3 \cite{brown2020language} to co-narrate stories. The AI system must track plot progression and character arcs while the human actors perform scenes. This event report details how a novel conversational agent was employed as creative partner with a team of professional improvisers to explore long-form spontaneous story narration in front of a live public audience. We introduced novel constraints on our language model to produce longer narrative text and tested the model in rehearsals with a team of professional improvisers. We then field tested the model with two live performances for public audiences as part of a live theatre festival in Europe. We surveyed audience members after each performance as well as performers to evaluate how well the AI performed in its role as narrator. Audiences and performers responded positively to AI narration and indicated preference for AI narration over AI characters within a scene. Performers also responded positively to AI narration and expressed enthusiasm for the creative and meaningful novel narrative directions introduced to the scenes. Our findings support improvisational theatre as a useful test-bed to explore how different language models can collaborate with humans in a variety of social contexts.

Improv Theatre and AI

The paper situates improvisational theatre as a platform for exploring creative collaboration between humans and AI language models. Unlike prior systems that primarily place AI in virtual environments or character roles, this study examines AI narration interpreted by live human actors.

Improv Theatre and AI

  • Improvisational theatre provides a test-bed for studying the creative potential and collaborative use of AI language models.Improv requires performers to accept and justify novel contributions while adapting to changing narrative circumstances.Source: Improv Theatre and AI · p. 1
  • Earlier collaborative storytelling systems mostly placed human players alongside digital avatars, with dAIrector an exception that informed live improv on stage..Source: Improv Theatre and AI · p. 1
  • The case study brings AI and human collaborators together on stage, with live actors expressing the resulting narrative through physical, emotional, and verbal performance..Source: Improv Theatre and AI · p. 1
  • Prior improv-focused agents generated dialogue as characters, whereas this study examines GPT-3 performing as a narrator.The narrator role shifts AI contributions from within-scene character speech to broader story direction.Source: Improv Theatre and AI · p. 2

Methods

The study combines GPT-3-based narration with real-time human curation, live performance controls, an avatar, and surveys of performers and audiences. The methods emphasize timing, filtering, and operator judgment as integral parts of the AI-human performance.

Datasets for AI improv

  • Large language models were used instead of collecting a new improv-specific dataset because their pretrained corpora provide broad thematic and linguistic diversity.The paper hypothesizes that fiction, synopses, and plot summaries are suitable sources for narrator training.Source: Datasets for AI improv · p. 2

Live curation, mitigating of model bias

  • Automated blocklists and Perspective API checks filtered offensive, inflammatory, hateful, or sexual content, while the operator could omit or reword outputs.This design places responsibility for selecting acceptable narration partly with the human curator.Source: Live curation, mitigating of model bias · p. 2

Interactive live AI narration on stage

  • Each operator-entered sentence was added to the scene context, after which GPT-3 generated three candidate sets of up to 100 characters for selective narration.The operator could choose none, one, or several suggestions in any order.Source: Interactive live AI narration on stage · p. 2
  • The operator typed prompts, selected suggestions, and could interrupt scenes while actors performed, preserving human control over timing and liveness.Timing was treated as an artistic and comedic decision made during the unfolding performance.Source: Interactive live AI narration on stage · p. 2

Story initiation

  • An automated novel-first-line prompting system was abandoned after rehearsals because its outputs were not informative enough for short scenes and actors preferred initiating stories themselves..Source: Story initiation · p. 2

Avatar for the AI narrator

  • The AI narrator was personified by a robot avatar controlled through the operator’s facial expressions, while a human voice provided expressive delivery.The operator stood behind the screen projecting the avatar on stage.Source: Avatar for the AI narrator · pp. 2–3

Evaluating AI in performance

  • The evaluation used a rehearsed 50-minute performance, two public performances, a partial transcript, and open-ended surveys of five performers and nine audience members.The study included a 12-minute AI-narrated long-form scene and followed approved institutional ethical standards.Source: Evaluating AI in performance · p. 3

Results

In the live performances, the operator selected GPT-3 suggestions that helped develop the improvised narrative while actors established characters and relationships. Audiences and performers generally responded positively, though some performers experienced the suggestions as random or inconsistent.

Results

  • GPT-3 generated 455 suggestions during a 12-minute scene between six actors, but the operator selected only a subset for presentation.The operator simultaneously observed the scene, entered context, voiced and animated the avatar, and controlled interventions.Source: Results · p. 3
  • The selected narration introduced characters, conflicts, transformations, and resolutions that advanced the story from a pizzeria setting to Sandra becoming a pastry chef.The episode included the burly man and his wife as mentors and concluded with the boss apologizing and rejoining the group.Source: Results · p. 3
  • AI interventions became more frequent after actors established characters and relationships, allowing the cast to form an initial narrative arc before heavier narration began..Source: Results · p. 3

Audience and performer response

  • Six of nine audience respondents reported satisfaction with AI narration, while two were dissatisfied and one was neutral; five most enjoyed the AI-narrated scene.Seven respondents identified the presence of AI as the strongest motivation for attending.Source: Audience and performer response · p. 3
  • All five performers were satisfied with the AI’s ability to move the story forward, although three slightly agreed that it mainly introduced absurd or random information..Source: Audience and performer response · pp. 3–4
  • Performers said the AI’s unexpected directions helped them develop character physicality, relationships, emotions, and plot while requiring them to justify surprising changes.One performer described the narration as advancing the plot without being overly prescriptive.Source: Audience and performer response · p. 4

Discussion

The AI narrator introduced meaningful plot developments, encouraged performers to adapt character status, reduced cognitive load, and helped transform errors into useful comic opportunities.

  • The AI introduced mentor characters and a classical transformation arc that generated novel, meaningful plot points with human timing decisions.Source: Discussion · p. 4
  • Inconsistent character motivations became narrated reversals that pushed performers to adapt status and produced clearer story beginnings, middles, and ends.Source: Discussion · p. 4
  • The narrator appeared to reduce improvisers’ cognitive load, letting them focus on relationships rather than simultaneously plotting and reacting.Source: Discussion · p. 4
  • Human errors that introduced seemingly random characters were accepted and reshaped into comic relief without disrupting the evolving narrative.Source: Discussion · p. 4

Conclusions

The study positions live improvisational theatre as a test-bed for human-AI creative collaboration while emphasizing human responsibility for interpreting and curating model output.

  • Improvisational theatre combines culture, social interaction, physical expression, and emotion, making it an AI collaboration test-bed that may uplift artists.Source: Conclusions · p. 4
  • Because language models mirror narrative tropes and biases, automated filtering and human selection retain responsibility for the story and its audience adaptation.Source: Conclusions · p. 4
  • The first live staging of an AI narrator co-creating improvised theatre identifies timing, interface usability, and imagined AI personality as future research factors.Source: Conclusions · p. 5
Loading 2109.14728v1…