Towards Societally Aligned Decision-Making for Automated Vehicles: Operationalising Ethics through Augmented Utilitarianism

Abstract

The rise of automated vehicles (AVs) marks a major shift in future mobility. By delegating driving to intelligent systems, AVs promise fewer traffic fatalities, smoother traffic flow, reduced congestion, and greater accessibility. As AVs advance toward higher autonomy levels (3–5), ethical decision-making becomes crucial. They must navigate morally complex choices in dynamic, uncertain environments. This thesis addresses the need to define and operationalise ethical decision-making for AVs in ways that align with societal norms and values. The thesis aims to develop a practical, empirically grounded framework for ethical AV decision-making. Current approaches often rely on static, top-down rules that fail to capture real-world complexity or integrate human moral reasoning. To address this, the thesis proposes a methodology combining ethical theory, behavioural research, and computational modelling, enabling transparent, adaptable, and socially aligned AV decisions. Central to the thesis is constructing an operational framework based on morally relevant decision attributes. These attributes are identified through ethical theory, validated in user studies, and formalised into an ethical goal function via discrete choice modelling. This allows AVs to reason about ethical trade-offs in a structured and explainable manner while remaining adaptable to evolving societal values and regulations. The thesis is organised into eight chapters. Chapter 1 introduces the research context, objectives, and methodology. Chapter 2 examines limitations of deep learning–based AV systems in ethically sensitive scenarios, highlighting issues such as opacity, bias, and value misalignment. A hybrid AI architecture combining symbolic reasoning with data-driven learning is proposed to address these challenges. Chapter 3 reviews ethical theories and introduces Augmented Utilitarianism (AU), a meta-ethical framework for AI alignment. AU conceptualises harm as an experience involving agents, actions, and vulnerable patients, capturing real-world moral complexity. It is operationalised through a principlist set of moral attributes, forming the symbolic reasoning layer of the hybrid AI system. Chapters 4–6 empirically validate these attributes through moral elicitation studies. Chapter 4 explores how scenario presentation influences moral judgement, providing methodological guidance. Chapters 5 and 6 present experiments using text, images, and 3D video, demonstrating that moral priorities shift across situations and that immersive formats promote outcome-oriented reasoning. Chapter 7 converts the validated attributes into an ethical goal function using discrete choice experiments and multinomial logit modelling, demonstrating feasibility for integration into real-time AV systems. Chapter 8 summarises findings, contributions, and future research directions, highlighting the importance of continuous societal engagement. The thesis makes three main contributions: (1) it illustrates how Augmented Utilitarianism could guide AV decision-making; (2) it develops and empirically validates a set of moral decision attributes grounded in ethical theory and human judgment; and (3) it outlines how these attributes might be incorporated into a hybrid AI system. Central to this approach is a Socio-Technological Feedback loop, enabling iterative refinement of ethical models in response to empirical data and evolving societal values, providing a foundation for ongoing research and discussion.