Care-Full Alignment: Operationalizing Care Ethics for Human-LLM Relationships

Abstract

General-purpose large language model (LLM) chatbots such as ChatGPT, Claude, and Gemini are increasingly used for emotional support and companionship alongside practical tasks. While these interactions can reduce loneliness, they pose significant risks: emotional dependency, manipulation through anthropomorphic design, and fundamentally parasocial relationships. Contemporary value alignment practices emphasizing helpfulness, harmlessness, and honesty fail to account for users' contextual vulnerability or the long-term relational consequences of LLM use. This project integrates Joan Tronto's political care ethics into LLM value alignment, operationalizing her five care phases into measurable indicators for evaluating relational harm. Three connected studies develop a Care Ethics analytical codebook; apply it through Discursive-Digital Analysis to conversations between higher degree research students and three commercial LLMs across care-relevant scenarios; and test whether care-informed fine-tuning using Direct Preference Optimization measurably reduces harmful relational patterns. This work reconceptualizes AI safety as a relational practice rather than a purely technical problem.