The Data-Production Dispositif

Machine learning (ML) depends on data to train and verify models. Very often, organizations outsource processes related work (i.e., generating annotating evaluating outputs) through business process outsourcing (BPO) companies crowdsourcing platforms. This paper investigates outsourced ML in Latin America by studying three platforms Venezuela a BPO Argentina. We lean the Foucauldian notion of dispositif define data-production as an ensemble discourses, actions, objects strategically disposed (re)produce power/knowledge relations labor. Our analysis comprises examination 210 instruction documents, 55 interviews with workers, managers, requesters, participant observation. findings show that discourses encoded instructions reproduce normalize worldviews requesters. Precarious working conditions economic dependency alienate making them obedient instructions. Furthermore, social contexts materialize artifacts, such interfaces performance metrics, limiting workers' agency normalizing specific ways interpreting data. conclude stressing importance counteracting fighting alienation precarization, empowering workers become assets quest for high-quality