Conference Agenda

Frontier AI developers discursively construct “ethical AI” through public-facing communications and governance documentation. This paper examines what meaningful distinctions emerge across organizations in their rhetorical framing, accountability structures, and positioning of institutional authority.

Specifically, we comparatively analyze public documentation from leading AI industry actors, including Anthropic, Google DeepMind, Microsoft, OpenAI, and other firms meeting defined frontier model criteria. We analyze how these organizations deploy and relate key vocabularies such as ethics, safety, alignment, risk, responsibility, governance and other normative and adjacent AI vocabularies across communication registers and over time.

Rather than assuming a single model of “AI ethics”, this study investigates whether firms retain ethics as a normative framework grounded in values and societal commitments, translate it into operationalized safety paradigms focused on technical risk mitigation, or combine these approaches in distinct ways to understand how those choices shape governance logics in practice.

This study asks: How do frontier AI firms discursively construct the relationship between ethics, safety, and governance vocabularies; and how do those constructions reshape notions of harm, accountability and authoritative expertise?

Methodologically, this study combines computational and qualitative analysis in a mixed-methods design.

We construct a longitudinal corpus of organization-authored public communications, including web-based news and research posts. Quantitative analysis includes structured keyword-concept frequency tracking to measure the prevalence and evolution over time of ethical, safety, and governance-related vocabularies (e.g. ethics, safety, alignment, governance, risk, responsibility) across organizations. We supplement this with unsupervised semantic modeling techniques (e.g. n-gram extraction, embeddings, density-based clustering, and dimensional visualization) to identify discursive pivots and thematic structures within and across firms.

Qualitative analysis extends beyond longitudinal communications to include key AI governance infrastructures that function as organizational “hubs”, such as model cards, AI constitutions, safety frameworks, and responsible AI portals or toolkits. These items, while often semi-static or “living” artifacts, provide insight into how organizations formalize and operationalize commitments related to ethics, safety, and governance in addition to longitudinal public-facing communications.

By comparing public communications with public governance infrastructure documents, we assess whether distinct registers emerge across contexts and how key terms, such as ethical, safety, and governance, are positioned relative to one another. This layered approach enables both large-scale comparative modeling and close interpretive analysis of how organizations articulate responsibility, harm, expertise, and institutional authority.

This research draws from information science, communication studies, science and technology studies (STS), and institutional governance scholarship. It treats organizational language as constitutive of ethical orientation, safety paradigms, and governance practice within frontier AI development through the lens of discourse theory and political economy of information (e.g., Howarth, 1998; Schiller, 2024). Drawing on discourse analysis and institutional theory, the study examines how accountability, legitimacy, epistemic authority, responsibility and harm are constructed through frontier AI developers' public documentation.

This study is novel in its systematic cross-organizational comparison of how frontier AI firms construct ethical AI across normative, procedural, and institutional domains, and in its inclusion of public governance “infrastructure” documents alongside longitudinal communications. While prior work has examined individual firms or mapped high-level ethics principles, few studies empirically compare how leading AI developers differentiate themselves rhetorically and institutionally, or how ethical, safety, and governance vocabularies vary across communication registers.

This research is directly relevant to contemporary technology and communications policy as governments worldwide develop AI regulatory regimes centered on risk classifications, audit mechanisms, disclosure mandates, and accountability standards. Public-facing discourse about AI ethics, safety, and governance increasingly functions as a signaling mechanism to regulators, investors, and civil society. Understanding how organizations construct and relate these AI vocabularies and narratives provide insight into how regulatory expectations may be shaped, narrowed, or expanded through industry discourse— with implications of whose expertise is legitimized, which harms are prioritized, and how oversight structures are designed.

This study will empirically map variation in how frontier AI development firms construct and relate ethics, safety and governance discursive terms across communicative domains and over time. The analysis will identify meaningful distinctions in how organizations define responsibility, articulate harm, and position institutional authority, and whether these vocabularies function as overlapping, hierarchical or distinct frameworks within and across firms' public documentation.

By systematically comparing these constructions across frontier developers, the paper provides an empirical foundation for understanding how ethical AI is being discursively institutionalized and how those constructions structure governance practice and public accountability.