Send message to Authors

Short abstract

Analyzing ground truth as a sociotechnical negotiation in Parkinson AI co-design, we show how lived experience is reduced to taxonomies and persona ontology. We propose reflexive ontology-making: a framework combining contestable classifications with continuous ontological adjustment.

Long abstract

This communication draws on the PD-TIPS.AI project, which develops an AI-powered conversational recommender system with and for people living with Parkinson's disease. In this context, the constitution of "ground truth" - essential for training and validating AI - does not rely on pre-existing datasets but emerges through an active co-design process.

Our objective is to demonstrate that ground truth is the result of situated sociotechnical negotiations. To do that, we analyze this process through two key artifacts:

a) The self-care taxonomy, which transforms the lived and embodied experiences of Parkinson's into structured categories. Through participatory workshops, we show that this process results not from medical consensus but from pragmatic compromises: a deliberate reduction of experiential complexity to create coherent training data.

b) The persona ontology, which distills patient diversity into archetypal profiles to personalize AI recommendations. These personas act as "boundary objects" facilitating dialogue between engineers, researchers, and patient-partners. However, they also create exclusions: patients with atypical trajectories become "edge cases", revealing the epistemic blind spots and the "foreclosed futures" of the modeling process.

This research proposes reflexive ontology-making as an alternative framework combining contestable classifications with continuous ontological adjustment. Rather than viewing AI brittleness as insufficient data, we reframe it as an epistemological challenge requiring systems designed for ongoing ontological maintenance. This approach offers a pathway toward more resilient AI—systems that accommodate revision as a structured practice rather than treating classifications as fixed foundations.