Artificiella Språkmodeller som Legitima Institutioner: En Studie kring Överföring av Symbolisk Makt

Abstract

This bachelor’s thesis is a study that aims to investigate whether large language models (LLMs) can be understood as institutions that reproduce and reinforce symbolic power. The theoretical framework draws upon Scott’s (2008) three pillars of institutions, Pierre Bourdieu’s (1991) definitions of symbolic power and habitus, George Lakoff & Mark Johnsons (2003) theory of conceptual metaphors and the mechanism of word embeddings. A theoretical analysis was conducted to examine whether LLMs, such as ChatGPT, exhibit institutional characteristics and contribute to the reproduction of social hierarchies. The analysis identifies patterns suggesting that LLMs reflect potentially amplify existing power structures through their data-driven outputs. In addition to addressing the main research question, the essay discusses the societal risks of letting LLMs be positioned as institutions, particularly regarding the symbolic influence and representation. The findings highlight the need for regulatory measures targeting data collection practices and promoting more equitable representation within training data för large language models.