Short abstract

Drawing on the PD-TIPS.AI project, we examines how "ground truth" for an LLM-based conversational recommender system (CRS) is progressively constituted through situated sociotechnical negotiation.

Long abstract

This communication examines how "ground truth" for an AI-powered conversational recommender system (CRS) is progressively constituted through situated sociotechnical negotiation. The present study draws on the PD-TIPS.AI project, a bilingual CRS supporting self-care for people living with Parkinson's disease, and presents six co-design iterations (2019-2026) through which a self-care taxonomy shifted from a clinician-centred classification to a patient-centred structure organized around five first-person intentions.

For this presentation, we will focus on participatory design workshops (using card-sorting method and dialogue-testing) involving patients and caregivers. The qualitative analysis revealed three main areas of tension relating to classification into the taxonomy: (1) clinical symptoms versus lived experience; (2) information versus intention to act; and (3) fixed categories versus boundary zones. Drawing on Mol's body multiple, Suchman's situated action, and Bowker and Star's classification theory, we trace how each tension was translated into a revised taxonomy, which includes merged categories, renamed terms, newly created domains, and cross-cutting tags that attempt to resolve the identified friction.

Since the taxonomy must consider lived experience not just clinical categories, we treat it as "infrastructure-in-the-making". This iterative process allows for the continuous revision and adaptation of categories and metadata via an interface that facilitates the ongoing adjustment of the taxonomy. Based on this, we propose a reflexive ontology-making approach. In other words, it is a framework for designing LLM-based CRS that grounds ontology construction in lived realities, real-world situations, and multiple contexts. The taxonomy illustrates this approach, paving the way for more "care-full" and resilient LLM-based CRS.