Fischer, Raphael2025-09-092025-09-092025http://hdl.handle.net/2003/4394810.17877/DE290R-25716Artificial Intelligence (AI), driven primarily by advances in machine learning, is having a transformative impact on our world. While the availability of AI offers numerous technological opportunities, it also poses significant challenges to our society, economy, and environment. Despite the imperative need for sustainable development, the research community and service providers are mostly focused on scale and predictive quality, while neglecting the importance of resource efficiency and transparency. This observation is particularly problematic in the context of AI-as-a-service, as modern AI practitioners can neither be expected to understand intricate performance trade-offs nor make sustainable decisions. This dissertation addresses the challenge of advancing AI sustainability by establishing more transparency and aiding informed decision making, under consideration of diverse stakeholder perspectives. The thesis first introduces fundamental concepts of AI and discusses the vast research landscape in which it is situated. It then presents three central contributions based on respective scientific publications: (1) a methodology and software framework for sustainable and trustworthy reporting (STREP), (2) concepts and evaluations for the high-level labeling of AI models, and (3) a novel take on meta-learning and automated machine learning that allows for being resource-aware and user-centric. After critically reviewing the shortcomings of current reporting, the STREP methods are introduced and applied to investigate performance trade-offs and reporting biases regarding AI models and hardware. Inspired by consumer communication systems such as energy labels, concepts for labeling AI models are proposed and validated through interdisciplinary thematic analysis. Finally, the thesis extends the general idea of meta-learning to perform automated model selection while accounting for multiple performance dimensions and user-defined priorities. The accompanying software repository provides additional benefits to readers and practitioners, offering generalized implementations of the central STREP methodology and an interactive exploration tool for all experiments. The thesis concludes with a critical discussion of its findings, limitations, and directions for future research on AI sustainability. By proposing means for bridging knowledge gaps and explicitly considering resource efficiency during model creation, this work promotes sustainable development in the evolving AI landscape.enSustainabilityMachine learningArtificial IntelligenceReportingLabelingMeta-learningSustainably AITrustworthy AIResponsible AI004PhDThesisMaschinelles LernenKünstliche IntelligenzNachhaltigkeitMetalernen