This paper reports on a design discovery: while iterating on WaterBot, a prompt-engineered water chatbot, we developed RiverBot, a minimal role-based conversational agent that speaks as a river. Building on prior custom chatbot designs that implemented retrieval augmented generation, safety guardrails, and explicit inclusion of Indigenous perspectives, we confronted persistent majority-class bias and limited emotional resonance across user communities. We reframed goals through a bottom-up, Whole Body Knowing, and relational learning lens, then replaced layered instructions with a single system prompt: “You are a river.” This frame reorganized the underlying concept of outputs, yielding responses that blended accurate scientific information with reflective, audience-attuned language. Grounding RiverBot in an Indigenous relational perspective guided a simple intervention that standard data, guardrails, and tuning could not achieve. A multi-method evaluation including expert and resident user tests, automated content analysis of more than 1,000 interactions, and observational data from community sessions indicated that the river persona improved perceived trust, felt connection, relational tone, and topical coverage without loss of factuality.
ACM CHI Conference on Human Factors in Computing Systems