Human-AI interactions are increasingly part of everyday life, yet the interpersonal dynamics that unfold during such exchanges remain underexplored.

In this talk we present findings from a large-scale citizen science installation, in which thousands of museum visitors co-authored stories with an LLM in a turn-taking setup. To disentangle the roles of human and LLM input, we contrast these field data with a simulated baseline where two LLMs completed the same task. Using sentiment analysis, semantic embeddings, and information-theoretic measures, we trace how alignment, exploration, and novelty emerge in collaborative storytelling. Our results reveal systematic differences between human and LLM inputs, highlighting the distinct ways in which human contributions shape narrative direction and creative divergence. Beyond specific findings, we introduce a methodological framework for analyzing dyadic interaction at scale, offering new tools for examining emotional alignment, semantic  exploration, and linguistic innovation in human–AI collaboration.