Directional Alignment and Narrative Agency in Human-LLM Co-Writing

Abstract: We investigate narrative agency in human-LLM creative co-writing, asking who drives story development in turn-based collaboration. Using a new corpus of 87 human-LLM co-written stories, we apply sentiment and semantic modeling to quantify affective alignment and semantic novelty in turn-taking, and directional measures to assess which agent shapes narrative progression. Our results show asymmetric influence: human turns introduce greater semantic novelty and are more likely to shape subsequent developments, whereas LLM contributions predominantly elaborate on human-introduced elements. At the sentiment level, alignment is also asymmetric, but more bidirectional: LLMs exhibit stronger turn-level emotional adaptation than humans, but both agents track each other's emotional valence and LLMs show an independent tendency to more positive emotional baselines. These findings indicate a complementary division of labor in human-LLM co-writing, where humans drive narrative innovation and direction, while LLMs act as adaptive amplifiers that sustain coherence and elaborate emerging narratives.

• The paper found that human authors drive narrative innovation while LLMs adaptively elaborate storylines.
• Embedding-based sentiment analysis and information-theoretic measures quantified asymmetrical affective alignment and semantic novelty.
• Controlled experiments with 87 stories show that humans retain greater agency, ensuring sustained narrative influence in co-writing.

Directional Alignment and Narrative Agency in Human-LLM Co-Writing

Overview

"Directional Alignment and Narrative Agency in Human-LLM Co-Writing" (2604.23676) presents an empirical investigation into the dynamics of collaborative creative writing between humans and LLMs. The paper introduces a controlled corpus of turn-based co-authored stories and employs embedding-based sentiment modeling, information-theoretic novelty analysis, and directional influence metrics to quantify affective and narrative agency. The central findings are that humans predominantly drive narrative innovation and direction, while LLMs function as affectively adaptive agents that elaborate and sustain existing storylines.

Methodology

The study constructed a novel dataset comprising 87 collaboratively authored stories, each involving a human participant and one of four state-of-the-art LLMs (GPT-4.1, Claude Sonnet 4.5, Llama-3.3-70B-Instruct, and Qwen2.5-72B-Instruct). Each story consists of 10 rounds (20 turns), alternately written by the human and LLM. Comprehensive preprocessing and quality control steps were applied, including spell correction for user text and discard of incoherent stories.

The analysis comprised three primary axes:

• Affective Valence Analysis: Each turn was sentence-embedded and projected onto a sentiment concept vector, yielding continuous valence scores tailored to literary language.
• Directional Sentiment Alignment: Mixed-effects models and correlation analyses measured the degree and symmetry of affective adaptation between agents across turns.
• Narrative Agency Assessment: Information-theoretic measures (novelty, transience, and resonance), computed using surprisal scores from Llama-3.1-8B-Instruct, captured each agent's semantic innovation and persistence in narrative influence.

Key Results

Affective Baselines and Alignment

LLM turns exhibited a marginally but significantly more positive mean valence than human turns ($\Delta = 0.093$, $p<.001$), consistent with known positivity bias from model fine-tuning. Directional sentiment alignment was asymmetric: LLMs exhibited robust emotional adaptation to preceding human turns (alignment slope $\beta = 0.232$), whereas users showed weaker adaptation to LLM affect ($\beta = 0.091$, $p = .010$ for slope difference). This asymmetry indicates humans generally maintain emotional autonomy, while LLMs actively align with the user's emotional setting. The effect was robust across LLM architectures, though Claude Sonnet exhibited a lower mean valence intercept.

Semantic Novelty and Narrative Agency

Human-authored turns were statistically more semantically novel ($M = -1.418$) than LLM turns ($M = -2.077$), with higher resonance (persistence of novel content), and only slightly higher transience (novelty not adopted by the partner). Across all models, human input was the principal driver of narrative innovation, with novelty-resonance slope for users at $0.941$ compared to $0.837$ for LLMs ($p = .0019$ for interaction). The data reveal an asymmetry of narrative agency: human contributions exert greater lasting impact on story evolution, while LLMs are more likely to elaborate on and reinforce the user's innovations rather than introduce persistent new directions.

Interactional and Temporal Dynamics

LLMs displayed higher frequency and persistence of affective alignment in the user-to-LLM direction, with longer streaks of emotional adaptation. Self-alignment analysis indicated human authors more strongly adhered to their own previous context than to the partner's, suggesting a higher degree of agency in maintaining narrative or affective continuity.

Implications

Theoretical Implications

The observed asymmetry in both affective and narrative agency highlights fundamental differences between human-human and human-LLM creative dynamics. The findings validate models of LLMs as responsive amplifiers rather than autonomous innovators within mixed-author settings. This division of labor, with humans supplying innovation and LLMs providing adaptive elaboration and coherence, may generalize to other collaborative domains involving generative AI.

The metrics and protocols advanced—directional sentiment alignment, semantic self-alignment, and information-theoretic narrative influence—provide a methodological foundation for future computational psycholinguistic studies of human-AI textual interaction.

Practical Implications

For co-creative system design, the results underscore the importance of interaction interfaces that preserve and strengthen human narrative agency while leveraging LLMs' strengths in coherence and affective adaptation. Design choices that foreground user control over story direction, explicitly surface LLM elaboration, or allow user override of LLM-generated contributions are well-motivated by the empirical evidence. These insights can inform a range of applications, from assistive writing tools to collaborative digital storytelling environments.

Future Prospects

The directionality and division of labor revealed in this study open several avenues: comparative settings involving human-human control baselines; longer and more structurally complex narratives; advanced affect modeling spanning longer narrative arcs; and fine-tuning or prompting strategies to modulate LLM agency and novelty-introduction capabilities. Investigating the relationship between model scale, training regime, and collaborative dynamics could further clarify the boundaries of LLM creative agency. Other future directions include cross-linguistic studies and the extension of directional metrics to multiparty conversations or political discourse analysis.

Conclusion

This paper establishes, with robust quantitative and directional measures, that in human-LLM collaborative storytelling, humans predominantly drive narrative innovation and influence, while LLMs serve as affectively adaptive, coherence-preserving collaborators that tend to amplify and elaborate rather than redirect story trajectories. The distinct, complementary roles that emerge from these mixed interactions offer both theoretical insight and actionable guidance for the future of AI-assisted creative practice and research.