Who Accommodates Whom? Bidirectional Linguistic Accommodation and Progressive Interpersonal Convergence in Human-AI...

Linguistic accommodation during human-AI interaction has been measured in only one direction at a time, leaving the relative magnitude of each side and the trajectory of within-conversation change unresolved. A symmetric within-versus-between conversation dissociation design applied to 1319 multi-turn English GPT-4o conversations from WildChat measures both user-side and model-side function word adaptation within the same data, revealing two distinct temporal dynamics. The model's adaptation is front-loaded, with strong initial accommodation at the first turn followed by stabilization, while users converge gradually across subsequent turns on interpersonal pronoun dimensions with no progressive change in topic-related categories. In 500 Switchboard human-human conversations, per-conversation similarity slopes are significantly negative (p=0.022), though the multilevel interaction is marginal (p=0.055). Because the pronoun dimensions on which users converge are the primary linguistic markers through which personality traits manifest in natural language use, this progressive convergence may represent a linguistic indicator of shifts in communicative self-presentation during extended human-AI conversation.