Artificial intelligence (AI) systems are increasingly used by governments across the globe to support or automate decision-making processes for the provision of fundamental public services. While AI offers an incredible potential for progress, an emerging body of literature highlights the challenges that AI-driven decision-making may pose for public sector ethics. A key aspect of these challenges is their association with different sources of epistemic opacity (algorithmic/legal/marketing campaigns/AI illiteracy) that may compromise the capacity of public institutions to adequately justify the use of AI systems, explain their outcomes, or challenge them and adopt corrective measures.

The primary goal of the thesis is to develop a normative theory framework for understanding the unique challenges to democratic decision-making that arise from the regulation and use of opaque AI systems in the public sector.

The thesis argues that AI’s epistemic opacity can lead to a dysfunctional cognitive environment that undermines citizens’ opportunities to act as co-authors of public decisions, either directly or through their authorized representatives. Insofar as AI systems obscure the comprehensibility and reviewability of public decision-making procedures, they undercut citizens’ capacity to engage in such processes and with their outcomes. Moreover, AI’s opacity undermines the traceability of authorship relations, thereby leaving undetected unauthorized uses of decisional power.

The thesis shed lights on the distinctive nature of algorithmic opacity as a normative challenge for the democratic regulation and public sector deployment of AI systems. It thus contributes to an emerging but still underdeveloped trend in normative political theory that study how AI-driven decision-making is reshaping the conceptualization and assessment of interactions between citizens and public institutions. Its theoretical framework can serve as a normative standard for AI policy assessments and as a background work to refine current studies in the fields of “AI fairness” and “Explainable AI”. In those contexts, it can contribute to better understand what normative principles should guide the adaptation of technological tools to meet distinctive democratic requirements for public decision-making. Additionally, the thesis develops a taxonomy of sources of epistemic opacity affecting AI systems, which contributes to their understanding as socio-technical artifacts, and is relevant for fostering a democratic politics of technology.