Comparison of Care - Ethical and FAT (Fairness, Accountability, Transparency)
Ethical Approaches in AI Translation
In this project, we critically examine why contemporary AI ethics is predominantly structured around fairness, accountability, and transparency (FAT), and whether these criteria are sufficient for evaluating AI systems in contexts that shape human relationships, vulnerability, and meaning. Drawing on Jonathan Cohn (2020) and care ethics (Gilligan 1996/ Held 2006), we argue that FAT-ethical approaches reflect an underlying “ethics of justice” that prioritizes formal equality, rule-based reasoning, and efficiency, while marginalizing dependence, relationality, and contextual judgment.
Focusing on AI-mediated language translation as an empirically tractable site at which these issues arise, we investigate how DeepL and ChatGPT instantiate FAT- versus care-ethical values. Through qualitative case studies (address, emotional vulnerability, gender) and quantitative LIWC analysis, we show that current systems tend to reproduce FAT-oriented patterns, often missing relational intent and reinforcing default assumptions (e.g.: gendered pronouns, unmarked formality choices, etc.). Care-ethical outcomes appear only sporadically and are likely unintended, stemming from broader training data rather than normative design.
Our findings suggest that a FAT-ethical dominance aligns not only with an ‘ethics of justice’, but reflects neoliberal hegemonies of optimization and efficiency, limiting AI’s capacity to address social and relational harms. In response, we propose a care-ethical AI-model card emphasizing voice, interdependence, contextual judgment, and vulnerability. Integrating care ethics into AI is not merely additive but requires revising the underlying concept of justice itself. At the same time, empirical and methodological limitations indicate that care-based AI remains an open and evolving research program.