Marc Hauer

• Rapid advancement and widespread adoption of Artificial Intelligence (AI) technologies have revolutionized numerous domains, transforming how we live, work, and interact. However, along with the immense potential of AI-based systems, there are significant challenges and risks that need to be addressed. One critical need is the regulation of such systems, particularly those making decisions about people as well as their property, so- called algorithmic decision-making (ADM) systems. Traditional regulatory approaches are inadequate for AI-based systems due to their vast input and state spaces, reliance on training data, complex inner structures, and the fast-paced nature of technological developments. Existing regulatory frameworks struggle to keep up with these novel challenges, requiring new approaches to ensure the responsible and ethical deployment of such technologies. Moreover, the current understanding and operationalization of key concepts such as fairness, explainability, and accountability in the context of AI-based systems remain vague, hindering the development of effective methods and solutions. Against this backdrop, the following scientific question arises: How can we address the challenges associated with the development, use, and control of AI-based systems to ensure their responsible and trustworthy deployment? This thesis employs a multidisciplinary approach to address the challenges associated with accountability in AI-based systems. The research is based on gaining an understanding of what such accountability means. Therefore, a generic software development process is dissected into sections, each examined separately to identify transparency and inspectability mechanisms. Building upon these mechanisms, various auditing procedures are explored, and the concept of certificates is introduced to ensure trustworthy audits. After this, testing of data-driven components and AI-based applications is described, focusing on fairness testing and its application within the audit procedures. As all approaches for promoting accountability require sufficient incentives to implement them, this thesis also reviews various approaches aimed at providing such incentives. It examines the role of the risk-based regulation approach suggested by the upcoming European AI Act and the recent Corporate Digital Responsibility (CDR) approach, highlighting potential benefits and areas for improvement. The multidisciplinary approach provides a comprehensive toolbox of methods and concepts to establish accountability in AI-based ADM systems. By providing insights into the effectiveness of these approaches, this work contributes to shaping future regulatory frameworks and promoting intrinsic motivations for accountability. Overall, this thesis not only identifies limitations in current approaches but also offers practical solutions and outlines future research directions to further advance the field of AI accountability. By fostering responsible AI development and usage, these results contribute to the long-term benefits of AI in society, ensuring that AI technologies can be deployed ethically, transparently, and in alignment with societal values.