Nathan Remillard (RAND Graduate School) has posted The Politics of AI Risk: Epistemic Authority in Large Language Models on SSRN. Here is the abstract:
Artificial intelligence (AI) risk debates are often polarized between concerns about catastrophic future harms and arguments that AI should be treated as a manageable “normal technology.” Though different, both of those perspectives assume that risk is an intrinsic part of AI systems and can be identified and mitigated primarily through technical evaluation and control. This article challenges that assumption by arguing that the most consequential risks of large language models (LLMs) are epistemic and institutional, rather than purely technical. Drawing on risk governance theory and Foucauldian conceptions of power-knowledge, the article conceptualizes LLMs as forms of epistemic infrastructure that shape how knowledge is produced, interpreted, and authorized within policy and governance processes. Under conditions of uncertainty and normative ambiguity, LLMs can stabilize dominant discourses and systematically marginalize alternative forms of knowledge. These dynamics introduce a distinct category of risk that cannot fully be addressed through existing technical AI governance tools such as auditing, evaluation, and monitoring. This contribution reframes AI risk as a problem of epistemic governance and highlights the limits of more technocratic approaches to managing social and institutional harms. The key argument is that recognizing such risks is essential for developing more reflexive and inclusive frameworks for AI governance.
Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum