#frAIday: Data quality for responsible medical AI

Data quality is one of the core goals of Responsible AI. In fact, in light of the EU AI Act achieving data quality becomes an important regulatory consideration in the development and deployment of algorithmic systems in high-risk domains, such as healthcare. The process of achieving appropriate data quality is a source of active debate. Computer Science scholarship and in particular Data Management Studies have previously defined data quality as constitutive yet contested in its dimensions, such as accuracy, structure, timeliness, and compliance. Nevertheless, little attention has been paid to how, under which conditions, and to what extent can data quality be achieved in high-risk domains.

In this talk, I will demonstrate how the experts’ “logics of practice” ground and define how and to what extent data quality dimensions are possible to be achieved within the increasingly demanding imperative to comply with data quality standards set by the EU AI ACT. This presentation is based on my extensive ethnographic study into the data practices of experts creating medical datasets for the development of diagnostic algorithmic systems in two health tech organisations in Western Europe. By unpacking how the data quality dimensions are contested and compromised in practice within organisations of different size and capacity, I argue that compliance emerges as a logic of practice that transverses and defines the rest of the data quality dimensions.