Towards Attuned AI: Integrating Care Ethics in Large Language Model Development and Alignment

Rayane El Masri, Aaron J Snoswell

Keywords: Ethics of Care, Value Alignment, Large Language Models, Normative Theories.

Abstract: How can the Ethics of Care (EoC) inform the development and value alignment of large language models (LLMs)? This paper proposes to investigate how a Care ethics framework, emphasizing relationality, attention to particularities, and contextual moral reasoning, can reshape existing approaches to aligning LLMs with human values. Mainstream AI alignment often draws on deontological or utilitarian principles, yet these frameworks can overlook the situated, affective, and power-sensitive aspects of moral life that Care ethics foregrounds. In this paper, we present two arguments for integrating Care Ethics into LLM development practices. First, we argue that LLMs often rely on overly generalized reasoning which contributes to various down-stream harms, including issues of bias. Second, we critique methods like RLHF and RLAIF for embedding narrow normative assumptions that neglect emotional and relational dimensions of human values. We argue that adapting LLM fine-tuning or alignment practices to incorporate Ethics of Care considerations may help address these issues, potentially laying the groundwork for better forms of LLM generalization and providing a pathway for more context-sensitive alignment of LLMs in care-relevant areas such as mental health, education, and social services.

Submission Number: 25