The Politics of Machine learning Evaluation From Lab to Industry

Artificial Intelligence (AI) applications are today implemented across various societal sectors, ranging from health care and security to taking part in shaping the media environment we encounter online. In the last decade there has been a significant shift in the field of AI, as the development of AI applications is no longer confined to the laboratory, but rather widely used and tested in and on societies. With this rapid industrialisation of AI, there is an increased need to understand the implications of both the development and deployment of these systems. While critical scholars have started to scrutinize different components of AI development, the study of evaluative practices in AI has received limited attention. A few studies have highlighted the importance of benchmarking practices and how these methods become integral in establishing the validity of the system and its success, which then enables widespread application. This paper presents a research agenda that outlines how to study machine-learning evaluation practices that move beyond the laboratory into industry applications and standardised validation practices. Based on emerging research and illustrative empirical examples from recent fieldwork, we argue to study machine-learning evaluation as a sociotechnical and political phenomenon that requires multi-level scrutiny. Therefore, we provide three analytical entry points for future research that address the political dynamics of (1) standardised validation infrastructures, (2) the circulation of evaluation methods and (3) the situated enactment of evaluation in practice.