Steinacker, LéaLéaSteinacker2023-04-132023-04-132022-02-21https://www.alexandria.unisg.ch/handle/20.500.14171/108985Advanced techniques in the field of Artificial Intelligence (AI) have been applied in commercial applications and public service across sectors to classify data, predict behaviors, and orchestrate choices. Today, experts agree that AI systems have immense economic, social, political, and environmental implications. But many recent institutional endeavors to assess them have been conceptually diffuse, overly focused on technical aspects at the cost of socialized context, and fueled by dichotomous narratives. Given the outsized influence of these sociotechnical systems, how can we capture the interdisciplinary factors that lead to their transformative effects on our social fabric? In this dissertation, I introduce my original notion of Code Capital, an interdisciplinary account of the intangible and material configurations that comprise an AI systems source of impact. Through the eponymous CODE framework, this new concept allows an analysis along four dimensions - Conception, Operations, Data, and Environment to express bespoke circumstances of each system, bringing to the fore its normative forces. To test the applicability of my approach, I conducted CODE analyses of two real-life AI systems using qualitative and quantitative techniques. For my first case study on facial recognition technologies, I present empirical results from a cross-country survey I conducted with a team of researchers that underline the need to contextualize AI systems in their social embedding. My subsequent CODE analysis of a particular deployed system illustrates the frameworks explanatory power for impact. I show how even thoughtful objectives risk producing unwanted outcomes and that the selection of material features has decisive effects on how the system is used. In my second case study on synthetic text-to-speech technologies, I examine the Code Capital of a system in its design phase to demonstrate how the concept can be used as a tool to guide the development process. My results show the importance of forecasting and contingency planning for potential misuse, such as the risk of identity fraud. Both case studies also emphasize the need for considering diverse representation in material design and training data to ensure inclusive participation and harm mitigation for users. Moreover, they demonstrate how centrally both the Conception and Environment dimensions contribute to the range of implications of a socially embedded AI system, which sets Code Capital apart from dominant existing approaches. Through the instructive CODE model, relevant stakeholders from the technological as well as the socio-political realm can employ a shared ontology to better anticipate and understand AI systems, with Code Capital as the novel descriptor of their potential power.enKünstliche IntelligenzMaschinelles LernenEthikEDIS-5201speech synthesisArtificial intelligenceAIimpact assessmentmachine learningsociotechnical systemsethicscapitalfacial recognitiondoctoral thesis