PUB - Publikationen an der Universität Bielefeld

This paper proposes a linguistic approach to trust in text-based conversational agents, such as LLM-based chatbots like ChatGPT. We argue that linguistic indicators of trust in human interaction should be tested to see if and how they can be applied to engage users in a trust-calibrated and user-agency enhancing interaction with LLM-based chatbots. Specifically, we consider conversational style, transparency about the status of LLM-based chatbots as next-word predictors, and human agency to be crucial aspects in the chatbot language design that promotes the calibration of trust in the capabilities, benevolence, and integrity of LLM-based chatbots. Furthermore, the proposed approach motivates users to think about the content, but also about how trust is expressed linguistically, and encourages them to think critically about the linguistic properties of chatbot responses.