The Linguistic Nature of Normative Judgment in AI Ethics: From Liberalism to Confucianism

Abstract

This doctoral thesis addresses a fundamental coordination problem in the governance of artificial intelligence: stakeholders across jurisdictions hold divergent normative judgments about AI technologies not primarily because they disagree about facts, but because they lack a shared theoretical orientation for connecting descriptive claims about AI to evaluative conclusions about what should be done. This thesis terms this condition the Cavemen-Luddites Problem – the predicament in which actors respond to transformative technology without adequate evaluative vocabulary, producing regulatory fragmentation, legal conflicts, and material losses.

To address this problem, the thesis takes normative judgment as its central unit of analysis. A normative judgment about AI comprises two elements: descriptive content (claims about what an AI system is and does) and evaluative content (claims about what ought to be done in response). The thesis is inspired by and refines the linguistic approach to philosophy of technology proposed by Mark Coeckelbergh, who argues that language and technology are deeply entangled. The unifying thematic arc of this research is the analogical proposition that normative judgments about AI are constrained and enabled by rules analogous to language rules, functioning in a manner analogous to linguistic expressions. While Coeckelbergh’s framework provides the foundational insight that normative reasoning about technology is a linguistic activity, this thesis identifies a specific limitation – the Composition Problem – concerning how descriptive and evaluative terms combine and shift in meaning within normative judgments. Drawing on Paul Grice’s distinction between semantics and pragmatics, the thesis demonstrates that evaluative terms in AI ethics (such as “fairness,” “privacy,” and “autonomy”) operate primarily as pragmatic implicatures: their normative force is context-dependent and cancellable, rather than fixed and semantically entailed. This finding is supported by empirical survey data from Hong Kong, which shows that respondents treat ethical values as near-equal universals in decontextualized settings but readily make trade-offs when presented with specific AI-application scenarios.

The thesis develops a Three-Feature Typology that organizes AI applications according to the dominant features of the data they process: physical-mathematical features (governed by deterministic laws, raising concerns of safety and liability), identity-constitutive features (linked to personal identity, raising concerns of privacy and dignity), and communicative features (involving meaning-making and strategic interaction, raising concerns of fairness and manipulation). The typology provides a sorting heuristic that predicts which evaluative vocabulary is structurally appropriate for a given AI application, while leaving the substantive deployment of that vocabulary underdetermined – to be supplied by the normative tradition operative in a given jurisdiction.

Two such traditions are examined in detail. Chapter 4 reconstructs Western AI governance through the lens of Isaiah Berlin’s distinction between negative and positive liberty. The institutional setting of United States AI governance – characterized by its reliance on corporate self-governance, contractual regulation, and reactive litigation in the absence of comprehensive federal legislation – is shown to be consistent with a negative-freedom orientation normatively underwritten by Nozickian libertarianism: autonomy as non-interference. The European Union’s risk-based regulatory framework – characterized by comprehensive risk classification, mandatory transparency, and preemptive compliance – is shown to be consistent with a positive-freedom orientation normatively underwritten by Rawlsian egalitarian liberalism: autonomy as the presence of enabling conditions for meaningful self-governance. The divergence between these approaches is diagnosed as a conceptual equivocation: both frameworks use the term “autonomy,” but invest it with different philosophical contents.

Chapter 5 introduces Classical Confucianism as a structurally contrastive evaluative framework. The chapter first establishes human relationships – codified in Mencius’s Five Relationships (wulun) – and their hierarchical relational obligations as the foundation of Confucian social order, and uses Ames and Rosemont’s concept of role ethics to demonstrate how this foundation differs structurally from liberal individualism. Through linguistic analysis of foundational texts, the chapter then reconstructs the Confucian concept of qi (器, vessels or artifacts) as the classical analogue to technology, and demonstrates that the ethical deployment of qi is semantically bound to correct Names (ming, 名) – the proper identification of social roles and their corresponding relational obligations. Unlike liberal frameworks that prioritize freedom and autonomy, Confucian ethics prioritizes the obligation of the state to provide for the material and moral well-being of the people within a hierarchical relational structure, yielding a proactive governance orientation grounded in relational duty rather than individual autonomy. In distributive justice, Confucianism supports appropriate distribution based on relational position rather than equal distribution irrespective of social role.

The thesis applies these frameworks to a comparative analysis of AI governance policies in the United States, the European Union, and China at two levels of analysis. At the macroscopic level, it surveys the primary policy judgments (or the broad governance frameworks) of each jurisdiction and evaluates the EU AI Act against the Three-Feature Typology. At the domain-specific level, it examines a secondary policy judgment: AI-mediated psychological counseling in education. Some states in the US restrict AI counseling to protect students from deception and unproven therapeutic claims in prioritizing negative freedom. In China, government guidelines permit and encourage AI counseling under supervision to address a severe shortage of human counselors, prioritizing relational obligation and consistent with the Confucian obligation of a benevolent government to promote people's welfare. The thesis demonstrates that both policy orientations are internally consistent with their respective normative traditions and that each tradition exposes genuine limitations in the other.

The thesis concludes by advancing Pragmatic Pluralism as the appropriate metaethical and methodological orientation for global AI governance. Pragmatic Pluralism combines the metaethical commitment of value pluralism – that multiple genuine values exist in irreducible conflict – with the pragmatic thesis that the normative weight of a value can be determined only by examining the full context of a choice situation. Drawing on Chang’s concept of covering values, the thesis argues that cross-jurisdictional policy conflicts can be rationally managed not by forcing convergence on foundational values, but by identifying shared covering values whose component parts receive different contextual weightings across jurisdictions. Global AI governance does not require a single moral language; it requires sufficient linguistic bridges to enable productive disagreement and coordinated action across normative traditions.

Date of Award	21 Apr 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Che Lan Linda LI (Supervisor) & Ho Mun CHAN (Supervisor)