Voices that connect: Exploring human-like attributes and eWOM in AI assistants

Large language model (LLM)-powered voice assistants rely on users’ electronic word-of-mouth (eWOM) to support their diffusion and ongoing improvement, yet little is known about which human-like attributes users are most likely to share with others. To address this gap, we employ a sequential mixed-methods design. Study 1 combines topic modeling with qualitative thematic analysis of posts from an online user community and identifies five attributes: general intelligence, personalized responses, human-like cues, warmth, and empathy. In Study 2, survey data from 334 voice assistant users are analyzed to test a media naturalness model linking these attributes to eWOM. Results show that all five attributes are positively associated with eWOM through automated social presence and user engagement. Automated social presence fully mediates the association between human-like cues and eWOM, while partly mediating the relationship for the other four attributes. By integrating machine-learning, qualitative, and survey-based methods, the research advances understanding of human-AI communication and provides insights for designing and managing voice assistant experiences that encourage positive eWOM.