Beyond bias: ethnographic approaches to “AI culture”

Abstract

Purpose This paper reconceptualizes biases in large language models (LLMs) from technical flaws requiring algorithmic fixes to intrinsic “AI cultures” amenable to ethnographic study. We argue that anthropological perspectives on knowledge and culture offer conceptual resources that enrich ongoing discussions of artificial intelligence (AI) bias, developing a framework for understanding how these systems participate in organizational meaning-making as culturally embedded actants. Design/methodology/approach Drawing on actor-network theory (ANT) and anthropological studies of knowledge production, we conceptualize LLMs as culturally embedded actants operating through visible networks of direct interactions and hidden networks of cultural assumptions. Ethnographic examples from organizational fieldwork with AI systems demonstrate how these networks shape workplace practices. Findings Rather than neutral computational processes, LLM responses emerge from complex cultural negotiations between users, organizational contexts and training data. Crucially, the cultural orientations of LLMs are not reducible to producer intentions: the model data assemblage generates emergent configurations that constitute a distinct analytical domain between the sociology of AI production and the study of AI use. The probabilistic mechanics of text generation inherently reproduce cultural patterns, making “unbiased” AI conceptually incoherent. Users who understand these dynamics can navigate between visible and hidden networks to reclaim agency. Research limitations/implications The study primarily focuses on the theoretical framework and illustrative examples from the authors' practices, which may limit the generalizability of findings. Future research should develop new forms of machine ethnography that treat model responses as cultural data, conduct longitudinal and comparative organizational ethnographies of AI integration and examine the emergent bridging figures whose ability to navigate hidden and visible networks may reshape organizational power dynamics. Practical implications Organizations should approach AI integration as cultural negotiation rather than technical deployment. Understanding LLMs as actants with their own cultures enables more reflexive and strategic AI adoption while revealing how expertise in AI interaction becomes a new source of organizational power. Social implications As AI-driven decisions are increasingly integrated into organizational life, our analysis suggests that expectations of removing bias rest on an incomplete understanding of what these systems actually do. Rather than neutralizing bias, chatbots introduce cultural configurations whose effects are distributed across hidden networks invisible to end users. As models grow more sophisticated, their cultural orientations will become more subtle, quietly reproducing particular assumptions about valid knowledge and rational decision-making in ways that demand critical scrutiny. Originality/value This paper introduces “AI cultures” as a theoretical concept bridging anthropological and organizational studies. Its distinctive contribution is the argument that LLMs generate emergent cultural orientations irreducible to their producers' intentions, constituting a third analytical domain that invites new forms of machine ethnography. This provides ethnographic methods for studying AI systems as cultural actors rather than neutral tools.

Ähnliche Arbeiten