The AI Shift Spectrum: A Sociotechnical Reconceptualization of AI Performance in Healthcare

Our work in progress interrogates the epistemological assumptions, and methodological limitations, of current conceptualizations of "shifts" and "drifts" in healthcare AI implementation. It advances a reconceptualization from a social studies of science and technology angle. By problematizing the artificial demarcation between technical, hard and social, soft determinants of algorithmic drift, we highlight how system performance emerges from complex socio-technical assemblages rather than isolated technical parameters. Our intervention yields three significant theoretical advancements: first, a comprehensive redefinition of "shift"; second, a multi-level taxonomy of shift sources spanning data infrastructures, organizational configurations, institutional arrangements, and socio-cultural dynamics; and third, a reconceptualization of AI development as an inherently socio-technical practice requiring interdisciplinary collaboration, mixed methods and capability building throughout the AI development lifecycle. This argument advocates for methodological pluralism that integrates causal modeling with qualitative social scientific approaches to address the full spectrum of performance-influencing factors across heterogeneous implementation contexts.