Several scholarly studies and journalistic investigations have found that automated decision-making in welfare systems burdens claimants by forecasting their behaviour, targeting them for sanctions and surveillance and punishing them without revealing the underlying mechanisms driving such decisions. This article develops an analytical framework combining three areas of concern regarding automation: how it might introduce surveillance and social sorting, how it can entail the loss of human discretion, and how it requires new systems of governance and due process. This framework steers investigations into whether and how automated decision-making welfare systems introduce new harms and burdens for claimants. A case study on automation processes in Germany’s unemployment benefit service’s IT system ALLEGRO applies this approach and finds that this system does allow for broad human discretion and avoids some forms of surveillance, such as risk-assessments from historic data, though it nevertheless increases surveillance of claimants through sharing data with external agencies. The developed framework also suggests that concerns raised in one area – whether loss of human discretion, surveillance, or lack of due process – can be mitigated by attending to the other two areas and urges researchers and policy-makers to attend to the mitigating or reinforcing factors of each concern.

In this paper we argue that qualitative longitudinal research (QLLR) is a crucial research method for studying automated decision-making (ADM) systems as complex, dynamic digital assemblages. QLLR provides invaluable insight into the lived experiences of users as data subjects of ADMs as well as into the broader digital assemblage in which these systems operate. To demonstrate the utility of this method, we draw on an ongoing, empirical study examining Universal Credit (UC), an automated social security payment used in the United Kingdom. UC is digital-by-default and uses a dynamic, means-testing payment system to determine the monthly amount of claim people are entitled to.We first provide a brief overview of the key epistemological challenges of studying ADMs before situating our study in relation to existing qualitative analyses of ADMs and their users, as well as qualitative longitudinal research. We highlight that, thus far, QLLR has been severely under-utilized in studying ADM systems. After a brief description of our study, aims and methodology, we present our findings illustrated through empirical cases that demonstrate the potential of QLLR in this area.Overall, we argue that QLLR provides a unique opportunity to gather information on ADMs, both over time and in real time. Capturing information real-time allows for more granular accounts and provides an opportunity for gathering in situ data on emotions and attitudes of users and data subjects. The ability to record qualitative data over time has the potential to capture dynamic trajectories, including the fluctuations and uncertainties comprising users' lived experiences. Through the personal accounts of data subjects, QLLR also gives researchers insight into how the emotional dimensions of users' interactions with ADMs shapes their actions responding to these systems.