Thinking Through Responsibilities in AI-Driven Healthcare

In this dissertation, I examine how responsibility should be understood in AI-driven healthcare. AI systems increasingly shape diagnosis, triage, treatment decisions, and access to care through predictive algorithms, digital health technologies (DHTs), and digital twins (DTs). These systems are often presented as improving efficiency, precision, and knowledge production within healthcare. At the same time, they reconfigure how care is delivered, how decisions are made, and how responsibility is distributed across socio-technical systems. The central concern of this dissertation is responsibility. When an algorithm misdiagnoses a patient, a DT generates a misleading simulation, or a DHT exacerbates inequities in low-resource settings, it becomes unclear where responsibility should be located, who should be held responsible, and who should act. Dominant retrospective, individualistic and attribution-focused (RIA) models locate responsibility in individual fault, intention, and control. These models assume discrete actors and linear causality. This dissertation shows that these assumptions do not hold in AI-driven healthcare. Harms arise through distributed processes shaped by data, design, and institutional structures rather than individual identifiable actions. The dissertation asks three questions: which conceptualisations of responsibility emerge in AI-driven healthcare, how useful they are, and which should be embraced. It focuses on key tensions and shows how different forms of responsibilities emerge, interact, and sometimes conflict. In contexts where decision-making is distributed and opaque, the question “who is responsible?” no longer suffices. Responsibilities must be understood as structural, differentiated, collective, forward-looking, epistemic, and relational. The dissertation consists of four chapters. Chapter One develops the conceptual foundation through Iris Marion Young’s social connection model. It shows that responsibilities are structural, differentiated, and forward-looking, grounded in actors’ positions and capacities within processes that produce harm. Chapter Two applies this conceptual foundation to DHTs in low-income countries. It examines how donor dependency, weak regulations, and resource limitations shape both benefits and harms. While DHTs may expand access to healthcare, they risk normalising lower standards, perpetuating data exploitation, and entrenching inequities. Responsibilities here are differentiated, with more powerful actors bearing proportionate obligations. Chapter Three examines DTs and develops the epistemic and relational dimensions of responsibility. It shows how data practices and modelling assumptions produce epistemic and relational harms. It develops a relational and epistemic account grounded in knowledge production, interdependence, participation, and shared moral labour. Chapter Four addresses responsibility gaps. It shows that these gaps reflect the limits of dominant frameworks rather than an absence of responsibility. Responsibility is displaced across socio-technical systems. Using an Ubuntu-inspired approach, it reframes responsibility as a collective and forward-looking practice focused on prevention and inclusion. Across the four chapters, this dissertation reconceptualises responsibility in AI-driven healthcare. It shows that responsibilities cannot be understood through RIA models alone. Instead, responsibilities are structural, differentiated, collective, forward-looking, epistemic, and relational. Responsibilities are formed and contested across networks of actors with different positions and capacities. The general conclusion is that responsibilities must be understood as a shared moral and institutional practice grounded in power, position, and interdependence, oriented toward preventing harm, sustaining justice, and ensuring accountability.