2026-10-13T10:47:46Z https://dcpapers.dublincore.org/oai

oai:dcpapers.dublincore.org:952683529 2026-09-10 dcmi-2026 openaire

Metadata-Conditioned Moral Grammars: How AI Companies Define Responsibility, Moral Language, Value Hierarchies, and Governance in AI Policy Documents Lee, Seul This study examines how leading AI companies, including Anthropic, Google, Meta, and OpenAI, construct ethical commitments in their governance documents, and how these normative constructions are systematically patterned by documentary and contextual metadata that situate each document within a broader governance architecture, including document type, governance level, declared purpose, scope, structural organization, section structure, organizational structure, intended audience, and temporal context. While prior work has increasingly treated AI governance documents as normative instruments influenced by institutional interests, sectoral differences, and political motivations, most of that work still tends to analyze these documents primarily in terms of themes, guiding principles, and policy discourse. This research extends that literature by arguing that AI ethics documents function as metadata-conditioned moral grammars: dynamic linguistic architectures that distribute responsibility, prioritize values, structure anticipatory governance, and guide decision-making under conditions of uncertainty. It investigates how AI governance documents produced by major technology corporations function as discursive infrastructures for the construction of ethical meaning. It analyzes how ethical frames including safety, innovation, autonomy, responsibility, and societal benefit are prioritized, sequenced, contextualized, and relationally embedded within these documents. Rather than understanding such principles as static normative commitments, this study examines their linguistic organization through a multidimensional analysis of lexical choice, grammatical structure, syntax, modality, agency, and semantic framing, asking how these features jointly constitute a metadata-conditioned moral grammar that distributes responsibility, structures anticipatory governance, and legitimizes institutional authority in AI governance. It examines how responsibility is both explicitly and implicitly attributed across various agents, including the organization, the AI system, users, developers, and abstract institutional entities, through patterns of grammatical agency and semantic positioning, while also identifying the linguistic mechanisms, such as passive constructions, hedging expressions, and conditional qualifiers that contribute to the diffusion or attenuation of responsibility. Dublin Core Metadata Initiative 2026-09-10 info:eu-repo/semantics/conferenceObject Text application/pdf text/html https://doi.org/10.23106/dcmi.952683529 https://dcpapers.dublincore.org/article/952683529 Dublin Core Metadata Initiative Conference Proceedings eng info:eu-repo/semantics/openAccess https://creativecommons.org/licenses/by/4.0/