Short Prompts, Long Replies: What Alignment Looks Like in Human–AI Dialogue

The purpose of this study is to examine how lexical and stylistic alignment appear in adjacent human–AI turns under strong role and length asymmetry in an instructional setting. Human–AI dialogue is often evaluated using conversation-level similarity measures, but these can hide how coordination is produced turn by turn. We analyzed written brainstorming dialogues from 88 students interacting with generative AI advisor avatars (967 turns). Using an R pipeline, turns were cleaned, lemmatized, and POS-tagged with UDPipe. Adjacent student–AI pairs were then scored for lexical overlap (Jaccard) and stylistic alignment (LSM). Lexical overlap was consistently low (content-lemma Jaccard ; bigram Jaccard ), consistent with limited opportunity for exact reuse when short student inputs are followed by longer AI responses. In contrast, LSM was moderate and variable (, ), and only moderately related to lexical overlap (), indicating that coordination can be expressed through framing even when topical words are not repeated. Both lexical and stylistic alignment were strongest early and declined over interactional time, suggesting reduced need for overt grounding once goals and roles stabilize. These findings show that low lexical overlap does not imply weak coordination and support turn-level, multi-dimensional analysis.