The Double-Edged Sword: Exploring Older Adults' Interaction and Imagination with an LLM-Enhanced Health Agent

The rise of generative AI, particularly large language models (LLMs), prompts a re-examination of human-agent interaction (HAI), inviting us to imagine novel roles for agents in society. This paper explores this theme of “Interaction and Imagination” through a multi-stage exploratory study on enhancing an embodied conversational agent (ECA) for older adults’ health monitoring. First, to ground our work in user needs, we conducted a co-creation workshop with older users (N=3) who had extensive, real-world experience with a pre-existing, intent-based health ECA. The goal was to identify its core limitations (limited contextual interpretation; rigid dialog) and to let users envision the key features of an ideal successor. This revealed a strong desire for an ECA with greater interactional flexibility, contextual understanding of health data, and social engagement. Second, guided by these co-created requirements based on voiced limitations and desires, we developed an LLM-enhanced prototype featuring a hybrid rule-based and generative dialog model. Finally, to evaluate whether this successor addressed the initial limitations, we conducted an exploratory within-subjects mixed-methods comparative study (N=7) directly contrasting the new LLM-ECA against the original intent-based system. Results indicate a ‘double-edged sword’: While the LLM-ECA was perceived as more ‘organic’, it was not rated higher on overall anthropomorphism and presented significant challenges in intuitiveness and cognitive load, exacerbated by user interface (UI) limitations such as push-to-talk voice activation. This study highlights the critical tension between the imagined potential of fluid, AI-driven conversation and the practical realities of interaction for older adults. We provide early empirical evidence underscoring the necessity of hybrid systems that balance generative adaptability with guided interaction, alongside robust UI design and structured onboarding. Our work offers actionable considerations for designing LLM-ECAs that effectively bridge the gap between imaginative possibilities and impactful HAI in sensitive domains.