Year: 2025
Type: dissertation
Abstract: As autonomous systems (AS) are increasingly adopted for high-impact applications, concerns arise about these systems making morally concerning decisions and performing harmful behaviours. These concerns have charged research efforts in value alignment, or designing and training AS to act in a way which is consistent with a certain set of values. Current human-centred solutions include training AS to i) prioritise the wellbeing of humans or ii) adhere to descriptive human ethics. This thesis investigates the prioritisation of human wellbeing as a normative grounding for AI ethics and value alignment. This thesis shows that anthropocentric accounts of moral value fail to motivate the claim that humans matter more than nonhumans. This thesis also examines the limitations of training AS to adhere to descriptive human values. Experimental results demonstrate that even when humans share moral intuitions, we struggle to translate these intuitions into a machine-readable format. Moreover, AS trained on human values often fail to align with relevant normative principles. This thesis proposes an alternative non-human-centred approach to value alignment. Ecological ethics is introduced and defended as an alternative normative grounding for value alignment, broadening the kinds of things afforded moral consideration to include nonhumans, such as plants and animals, and entire ecosystems. This thesis also presents a novel technical contribution, `modelled ethics', where domain-specific quantitative models are utilized to capture normative ethics. Experimental results show that modelled ethics value alignment outperforms human-set reward functions in adhering to a both human-centred and ecological ethics. This research de-centres the human in value alignment, offering a more inclusive AI ethics paradigm and creating more ethically robust AS.
Author: Rigley, Eryn
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7077307380