AI ethics unwrapped: understanding and shaping ethical AI through value sensitive design

This dissertation examines the real-world effects of artificial intelligence (AI) system design and explores how AI systems can be designed to achieve ethical outcomes. It focuses on the social, technical, and institutional dimensions of design, grounded in the understanding that AI is not merely a computational artifact but a sociotechnical system. As AI systems increasingly shape decision-making in high-stakes domains such as healthcare and hiring, concerns around fairness, transparency, and inclusion move to the forefront. Instead of retrofitting alignment with these values, this dissertation explores how Value Sensitive Design can support the proactive integration of ethical principles throughout the AI lifecycle, while also identifying where current approaches fall short.

The work is structured around five research questions (RQs), divided into two parts. Part 1 is based on two publications and focuses on understanding the effects of AI design, specifically on decision-making and inclusion. For addressing RQ1: How do design choices in algorithmic decision-making systems influence transparency and explainability of AI systems?, formal modeling is used to simulate ethical trade-offs in healthcare. The findings demonstrate that by making the decision architecture of AI systems explicit, we can generate measurable ethical impacts that can support us in designing systems that align with our values. RQ2: How does the design of AI-systems affect persons with disabilities, and what ethical, legal, and technical shortcomings contribute to these outcomes?, is explored through an interdisciplinary analysis combining ethics, legal evaluation, and technical assessments. The study reveals that current AI-based hiring tools often neglect the needs of persons with disabilities due to their design, legislation, and technical functionality. The findings point to the necessity of inclusive design practices that extend beyond narrow bias mitigation.

Part 2 is also based on two publications and one additional paper currently under review for publication, and turns to the conditions that shape ethical AI design, structured in three RQs. In response to RQ3: How do emergency contexts influence the development, ethical evaluation, and perceived trustworthiness of AI systems?, the dissertation presents a focus group-based case study on an AI-assisted diagnostic system. The study shows how urgency, uncertainty, and resource scarcity shaped both the technical design and its ethical framing, revealing how perceived trustworthiness of an AI system is co-constructed by contextual and organizational factors. RQ4: How do AI developers discuss and apply ethical principles in the process of designing prototypes of AI-based systems? is examined using the qualitative methods of design thinking workshops and interviews. The results demonstrate that developers engage with ethical values, but in context-specific ways that are influenced by individual motivations, and the best interests of users. Finally, RQ5: What is the current state of applying VSD to AI as reported in research studies, and which approaches have been proposed in the literature to address challenges for this application? is addressed through a systematic literature review of VSD applications to AI. The analysis reveals a lack of consistency in stakeholder identification, ambiguity in value translation processes, and limited documentation of iterative practices. While VSD provides a flexible framework for integrating ethics into AI development, its application can and must be refined. The dissertation proposes such refinements, including structured stakeholder identification, context-sensitive value translation, and iterative design.

These contributions advance our understanding of how design decisions of AI systems impact individuals and social groups, and how design decisions of AI systems are influenced by context, humans, and methodologies. This dissertation demonstrates that ethical impact is never incidental but embedded in AI architectures, organizational routines, and developer practices. By demonstrating this, it also shows that ethical AI development requires not only methodological rigor but also institutional support and awareness for the importance of design. Future research can build on these findings by defining ethical maintenance of AI systems post-deployment, refining ethical design methodologies such as VSD, and establishing social practices for AI developers.