Authors: Huon Curtis*, University of Sydney
Topics: Economic Geography
Keywords: Amazon Mechanical Turk, hidden labour, machine learning,
Session Type: Paper
Presentation File: No File Uploaded
High quality machine learning relies on accurately labeled training data. Data labelling is a task that relies on complex, geographically extended, labour networks that include labour platforms, such as Amazon Mechanical Turk (AMT), “data factories” located in emerging markets including China and India, and prison labor in developed economies. Workers on AMT lack basic workplace rights around collective bargaining (unionization), channels for redress of employer wrongdoing, and have no guarantee of payment for work properly performed. Further, facial recognition ML are often trained on data archives of photos of immigrants, child victims of crime and prisoners (foregrounding other hidden processes of generation and collection of data). Academic labour markets are also implicated in these labour networks as they are the largest user of labelling services such as AMT for their academic studies. This paper explores the hidden labour and processes of artificial intelligence. Much academic commentary on AI has the front-end, consumer-focused consent challenges (e.g. use of personal information in unexpected ways, the unpredictability and opacity of AI systems). Despite headlines “predicting” the risk of automation of occupations and tasks, the extent and type of human labour involved in the development of AI is mostly hidden, and this is relatively unaccounted for in existing literature. Data-labelling is a component of long-term trends, particularly outsourcing of mid-skilled routine cognitive labour. The process of labelling data is largely hidden in discussions of the social and employment effects of ML and AI.