Scandals in which governmental ADM tools played a role, have recently brought about political and societal debate about the potential harms to citizens that such automated systems potentially bring. This thesis focuses on ADM systems which contain an AI component. Based on the knowledge gaps perceived, the main research question for this study is: In public sector AI systems, what are emerging vulnerabilities for citizens and how do these translate into governance requirements for decision-makers? The approach to answer this question is an adjusted form of Theory Building from Case Study. A layered ’onion model’ presents four relevant contexts to consider for AI System vulnerabilities specifically in the public sector: AI model, Model deployment, Political-administrative, and Societal. The case study results demonstrate that dealing with vulnerabilities in one of the four model contexts often complicates dealing with vulnerabilities in the other contexts. Hence, the vulnerabilities model points to so-called governance requirements dilemmas.

The thesis explores the complexities and risks associated with Automated Decision-Making (ADM) systems in social welfare, focusing on the phenomenon of ”digital cages.” The digital cage concept refers to a situation where rigid systems, perpetuated through algorithmic misclassification, inadvertently trap citizens in bureaucratic complications without recourse. This thesis analyzes the digital cages concept from a system safety perspective, on the example of the Dutch Toeslagenaffair. The thesis centers on understanding how digital cages form within social welfare systems and seeks methods to mitigate their emergence through targeted interventions using system safety theory. The pivotal question addressed is: ”What are safety control interventions to curb the emergence of algorithmic decision-making systems-induced digital cages in the context of national social welfare administration in the Netherlands?” The research concludes by highlighting the potential of assumption-based leading indicators. Assumptionbased leading indicators are predictive measures that rely on underlying assumptions about how certain inputs or actions correlate with future outcomes. The leading indicators provide a novel approach to enhancing system safety in social welfare administration by addressing the underlying assumptions of the system in operation. This offers a substantial potential to transform how social welfare systems manage and utilize ADM, potentially reducing the incidence of digital cages. By detecting early signs of hazardous systems states digital cages can be mitigated before they solidify into systemic issues. The thesis not only addresses the immediate concerns regarding digital cages but also opens pathways for future inquiries into safer and more equitable administrative practices. Future research is suggested to further refine these indicators and explore their applicability in diverse administrative contexts, ensuring they can effectively adapt to the evolving landscape of public administration and technology.