Abstract This paper argues that bias in large language models (LLMs) is not a technical malfunction but a structural effect of the epistemic and institutional regimes in which these systems are developed and aligned. Drawing on Michel Foucault’s concept of power/knowledge and postcolonial theory, it conceptualizes LLMs as productive discursive apparatuses that normalize particular ways of knowing, speaking, and reasoning. Unlike earlier classificatory systems associated with surveillance capitalism, generative models exercise power primarily at the level of discourse: they shape how explanations are structured, how problems are framed, and what counts as reasonable or legitimate articulation. Particular attention is given to Reinforcement Learning from Human Feedback (RLHF), which translates situated human judgments of “helpfulness” and “appropriateness” into scalable optimization objectives. Through this process, historically contingent norms are transformed into algorithmically stabilized standards, producing truth effects without corresponding truth procedures. Rather than framing bias as an anomaly to be corrected through technical refinement, this paper advances a critical epistemology of AI that foregrounds normalization, subject formation, and the reorganization of regimes of truth. By situating LLMs within broader technopolitical and epistemic structures, the analysis shifts the debate from fairness metrics toward the deeper question of how generative AI participates in shaping the horizons of intelligibility and the conditions under which knowledge becomes authoritative.
Theodoros Kouros (2026) studied this question.