What is Safety? Corporate Discourse, Power, and the Politics of Generative AI Safety

Paper Title

What is Safety? Corporate Discourse, Power, and the Politics of Generative AI Safety

Publication Info

• Topic area: Critical analysis of corporate discourse on generative AI safety.
• Keywords: AI safety, corporate discourse, generative AI, accountability, governance, sociotechnical systems, critical discourse analysis, power dynamics, AI literacy, participatory governance.

Background and Problem

• Problem / challenge: The concept of "safety" in generative AI is predominantly defined and communicated by corporate actors, often serving their interests. This framing risks reproducing corporate priorities, limiting alternative governance approaches, and obscuring power dynamics.
• Significance: Understanding how safety is constructed in corporate discourse is crucial for addressing societal impacts, ensuring accountability, and promoting equitable AI governance.
• Motivation and related work: Prior research has explored AI safety risks, mitigation strategies, and sociotechnical dimensions. However, little attention has been paid to how corporate narratives shape public understanding, research priorities, and governance frameworks. This paper focuses on analyzing corporate safety discourse to reveal embedded power relations and political implications.

Solution

• Proposed approach: Critical discourse analysis (CDA) of safety-related public documents from three major generative AI companies: OpenAI, Google, and Anthropic.
• Novelty:
  1. Situates "safety" as a sociotechnical discourse requiring critical examination.
  2. Analyzes how corporate actors construct authority, diffuse accountability, and frame risks.
  3. Highlights the political implications of corporate safety narratives for AI governance.
  4. Proposes a critical-discursive dimension to AI literacy and governance.
• Procedure and key techniques:
  • Selection of 75 safety-related documents from OpenAI, Google, and Anthropic.
  • Use of critical discourse analysis to examine structural, detailed, and synoptic dimensions of safety discourse.
  • Identification of themes such as responsibility, governance, risk mitigation, authority construction, and rhetorical strategies (e.g., metaphors, global framing).

Results

• Concrete findings:
  • Responsibility is framed as distributed across companies, users, and governments, but accountability for harms remains vague.
  • Governance emphasizes internal structures and selective regulatory collaboration, with calls for "surgical" regulations to avoid stifling innovation.
  • Risks are framed as broad and evolving, requiring iterative mitigation practices like red-teaming and user feedback.
  • Authority is constructed through technical expertise, ethical positioning, and institutional arrangements.
  • Safety is portrayed as dynamic and globally significant, with metaphors drawn from high-risk domains (e.g., nuclear power, aviation).
• Advantage over baselines: Provides a critical lens on corporate safety narratives, revealing how they shape public understanding and governance while obscuring power dynamics and accountability gaps.
• Experiments / evaluation:
  • Analysis of 75 public documents using CDA.
  • Identification of recurring themes and rhetorical strategies across three companies.
  • Comparison of corporate narratives with broader sociotechnical and governance frameworks.
• Limitations and future work:
  • Focus on public documents excludes internal practices and informal discussions.
  • Limited to three major companies; broader datasets could reveal additional patterns.
  • Interpretive nature of CDA; complementary methods (e.g., surveys, experiments) could assess audience perceptions.
  • Need for longitudinal studies to track evolving corporate discourse.

Summary

This paper critically examines how leading generative AI companies construct and communicate the concept of safety through public-facing documents. Using critical discourse analysis, it reveals how companies frame responsibility, governance, and risks while constructing authority and diffusing accountability. The findings highlight the political implications of corporate safety narratives, showing how they shape public understanding and governance while limiting external oversight. The paper contributes to AI literacy by proposing a critical-discursive dimension and informs AI governance by exposing the performative nature of participatory rhetoric and corporate framing.