328
Alignment, Exploration, and Novelty in Human-AI Interaction
arXiv:2512.17117v1 Announce Type: new
Abstract: Human-AI interactions are increasingly part of everyday life, yet the interpersonal dynamics that unfold during such exchanges remain underexplored. This study investigates how emotional alignment, semantic exploration, and linguistic innovation emerge within a collaborative storytelling paradigm that paired human participants with a large language model (LLM) in a turn-taking setup. Over nine days, more than 3,000 museum visitors contributed to 27 evolving narratives, co-authored with an LLM in a naturalistic, public installation. To isolate the dynamics specific to human involvement, we compared the resulting dataset with a simulated baseline where two LLMs completed the same task. Using sentiment analysis, semantic embeddings, and information-theoretic measures of novelty and resonance, we trace how humans and models co-construct stories over time. Our results reveal that affective alignment is primarily driven by the model, with limited mutual convergence in human-AI interaction. At the same time, human participants explored a broader semantic space and introduced more novel, narratively influential contributions. These patterns were significantly reduced in the simulated AI-AI condition. Together, these findings highlight the unique role of human input in shaping narrative direction and creative divergence in co-authored texts. The methods developed here provide a scalable framework for analysing dyadic interaction and offer a new lens on creativity, emotional dynamics, and semantic coordination in human-AI collaboration.
Abstract: Human-AI interactions are increasingly part of everyday life, yet the interpersonal dynamics that unfold during such exchanges remain underexplored. This study investigates how emotional alignment, semantic exploration, and linguistic innovation emerge within a collaborative storytelling paradigm that paired human participants with a large language model (LLM) in a turn-taking setup. Over nine days, more than 3,000 museum visitors contributed to 27 evolving narratives, co-authored with an LLM in a naturalistic, public installation. To isolate the dynamics specific to human involvement, we compared the resulting dataset with a simulated baseline where two LLMs completed the same task. Using sentiment analysis, semantic embeddings, and information-theoretic measures of novelty and resonance, we trace how humans and models co-construct stories over time. Our results reveal that affective alignment is primarily driven by the model, with limited mutual convergence in human-AI interaction. At the same time, human participants explored a broader semantic space and introduced more novel, narratively influential contributions. These patterns were significantly reduced in the simulated AI-AI condition. Together, these findings highlight the unique role of human input in shaping narrative direction and creative divergence in co-authored texts. The methods developed here provide a scalable framework for analysing dyadic interaction and offer a new lens on creativity, emotional dynamics, and semantic coordination in human-AI collaboration.
Anonymous
No comments yet.