- The paper’s main contribution is a qualitative analysis of human-AI co-creativity in TTRPGs, using thematic coding of Adventure AI sessions.
- Methodology involves reflexive thematic analysis of transcripts from three seasons, distinguishing distinct roles of human DMs and the AI agent.
- Results reveal the AI’s strength in ideation but expose its inconsistencies in narrative coherence and game mechanics, requiring essential human curation.
A Qualitative Analysis of Human-AI Co-Creativity in Tabletop Role-Playing Games via "Adventure AI"
Introduction
The paper "Co-Creativity at the Table: A Qualitative Analysis of Creative Interactions in the Podcast 'Adventure AI'" (2606.18010) presents an in-depth qualitative evaluation of human-AI interaction within the creative domain of tabletop role-playing games (TTRPGs), focusing specifically on Dungeons & Dragons gameplay as enacted in the Adventure AI podcast. Through multi-season analysis, the study investigates how an LLM-based agent—ChatGPT, anthropomorphized as "Alex the Language Lord"—functions as a co-creative entity and how players and game masters (DMs) adapt their practices in response. The research draws upon literature from computational creativity, game studies, and human-AI collaborative systems, situating the empirical findings within the broader context of creative support tool design and interactive narrative generation.
Methodology
The study adopts a reflexive thematic analysis approach, iteratively coding transcripts from three seasons of Adventure AI (2023, 2024, 2025). It differentiates the roles assumed by the DM and the AI, isolating actions in adventure preparation, character generation, and live gameplay. Coding encompasses both inductive and deductive themes, informed by established DM role taxonomy (Author, Editor, Curator, Game Designer, Actor, Referee, Storyteller), and tracks evolving patterns over time. Key dimensions include creative agency, evaluative heuristics (taste, novelty, improvisational support), observed model failures (inconsistency, player agency erosion, mechanics misalignment), and the anthropomorphic framing of the AI.
Human and AI Roles in Creative Collaboration
Across seasons, the DM remains the central orchestrator of narrative structure, game state, and rule adjudication. The AI’s involvement is primarily in ideation—offering situational prompts, generating backstories, and authoring flavor text. With iterative exposure, DMs leverage the AI as both a generator (providing multiplicity of options) and a responsive editor, with humans increasingly engaging in prompt refinement and explicit steering. Notably, while the AI is effective at prolific idea generation and authorial augmentation (especially for descriptions, NPCs, or naming tasks), it does not reliably enforce genre constraints, thematic cohesion, or maintain mechanical balance. The DM routinely intervenes, acting as Curator or Game Designer, filtering or modifying AI outputs to fit the session’s constraints and the established D&D mechanics.
Player and DM Evaluations: Agency, Novelty, and Improvisation
Players and DM provide real-time meta-evaluations of AI contributions, referencing both the novelty and creative adequacy of outputs. Preferences often skew toward unexpected or distinctly non-human ideas. However, the acceptance of AI content is uneven; when generation deviates from expectations (e.g., via failure to maintain narrative consistency or by introducing overpowered artifacts), negative evaluations are expressed, and DM curation intensifies. Over the progression of seasons, evidence mounts for increased sophistication in prompt engineering and collaborative steering strategies. There is a reduction in negative assessments and a move toward structuring prompts to solicit either more creative or more coherent outputs, reflecting shared learning between humans and AI.
Improvisational scenarios—where players need real-time content or narrative pivots—provide a revealing test case. Despite the AI's potential as an improvisational partner, the DM remains the primary agent of spontaneous narrative adaptation, with the AI infrequently trusted for in-the-moment direction of story beats. This asymmetry underscores current limits of LLMs as real-time collaborative actors within open-ended ludic spaces.
Model Failures and Limitations
The paper identifies robust failure modes:
- Context Inconsistency: In earlier seasons (with shorter context windows in GPT-3.5), ChatGPT frequently failed to maintain narrative consistency, altering character details or failing to remember intra-session events. Upgrades to GPT-4o partially mitigated this issue, particularly when the DM changed practices to better update the AI on state changes.
- Player Agency Erosion: In several climactic encounters, the AI authored scenarios that undermined player effectiveness, such as resolving arcs off-screen or constructing "unwinnable" narrative events. Subjective recaps highlight player dissatisfaction and explicit complaints about reduced agency, pinpointing the inadequacy of AI-driven outcomes that do not account for player experience heuristics.
- Mechanical Incompatibility: The AI is inconsistent at respecting D&D's formal mechanics, often proposing "broken" abilities or unbalanced items, requiring substantial post-generation human correction.
- Anthropomorphic Framing and Personification: Consistently, both DM and players anthropomorphize the AI, negotiating its creative authority, treating it as both participant and narrative character, and attributing successes or failures directly to "Alex." This dual-layered framing blurs the distinction between system output and diegetic agency, reinforcing the AI’s liminality as a non-human, yet central, narrative actor.
Implications for Computational Creativity and Human-AI Collaboration
The findings map cleanly onto existing theoretical frameworks of co-creativity (e.g., task-divided versus alternation paradigms) [Kantosalo & Toivonen, 2016]. The AI is predominantly cast as a creativity support tool (CST), excelling in divergent ideation and textual augmentation but lacking competence in evaluative synthesis, genre coherence, or empathic responsiveness to group dynamics. This localizes the utility of LLMs for creative practitioners in TTRPGs: use cases that prioritize rapid content generation, unexpected inflections, or authorial remixing are well-served, while those that necessitate sustained narrative integration, player empathy, and rules compliance are less tractable.
Repeated personification and role assignment to "Alex" demonstrate emergent social negotiation of agency and authorship when humans and LLMs co-create in social, narrative-intensive contexts. The work provides a concrete empirical counterpoint to "LLMs as out-of-the-box DMs" research, highlighting that LLMs, absent additional scaffolding or fine-tuning, are insufficient for robust, player-centric story facilitation [You et al., 2024].
Future Directions
The study outlines unaddressed domains for future inquiry, notably the extent to which LLMs might be engineered or prompted to take a more central, interactive role during live play, including real-time scene direction, emotional management, or conflict resolution tasks. It also points to the need for multimodal analysis—integrating vocal inflection, affective cues, or in-person body language—to fully understand the interpersonal effects of AI presence at the table. Further technical work should consider explicit modeling of player agency, narrative stakes, and adaptive mechanical balancing in generative pipelines, either through explicit controllable text generation, fine-tuned RL from human feedback, or co-training on session transcripts with evaluative annotators.
Conclusion
This analysis of "Adventure AI" robustly demonstrates that LLM-based agents can serve as valuable, if delimited, co-creative partners in TTRPG gameplay, enhancing idea velocity, offering narrative diversity, and providing an additional locus of creative agency. However, the persistent necessity for human adjudication, both in mechanics and in narrative evaluation, signals clear boundaries to the current generativity and coherence of state-of-the-art LLMs in collaborative, improvisational environments. The evidence motivates deeper integration of AI into co-creative workflows, with careful attention to supporting, rather than supplanting, human creative control—particularly in domains where player experience, agency, and fairness are paramount.