Conversational Lexicons in Human–AI Interaction: A Mixed-Methods Study of Generative AI–Supported Lexical Decision-Making

Large language models are increasingly used as interactive language tools, yet their effects on fine-grained human–AI decision-making remain underexamined. This mixed-methods experiment investigated a ‘conversational lexicon’ workflow in which student translators used a generative AI system to explore lexical alternatives while retaining responsibility for final decisions. Sixty senior Arabic-speaking student translators working from English or French into Arabic were randomized within language cohort to a GenAI-supported condition (n = 30) or a traditional digital-resource condition (n = 30). Outcomes included final translation quality, perceived cognitive load, creative confidence, pre-task self-reported GenAI-use functions, screen/audio records, and interviews. The GenAI group obtained higher quality scores than the control group (M = 46.47 vs. 35.40), with Welch’s t(44.62) = 51.19, p < .001. It also reported lower cognitive load and higher creative confidence; a joint multivariate test was significant, F(2, 57) = 554.24, p < .001. Pre-task responses showed that contextual explanation and alternative phrasing were the most frequently reported GenAI uses. Qualitative evidence highlighted iterative prompting, cross-checking, contextual judgment, and concerns about over-reliance. The findings support a human-centered interaction account in which natural-language interfaces expand the option space for lexical decisions but require active verification. The unusually large between-group effects should be interpreted cautiously because outcomes showed restricted within-group dispersion and arose from a tightly controlled, task-specific intervention.