Large language models adapt the political unconscious and repress heteroglossia

Abstract This article examines the ideological function of large language models (LLMs) by analysing how they adapt existing texts when asked to produce fictional narratives. Rather than functioning as a form of Bakhtinian heteroglossia, a synthesis of multiple, diverse, and centrifugal voices, the ­article argues that LLM discursivity instead operates as part of a centripetal, ideological project that expresses the ideas of the ruling class. The study uses examples from a series of prompts given to Gemini, a Google-developed LLM, to generate stories about social problems such as poverty, revolution, and environmental disaster. The findings demonstrate that Gemini ­consistently adapted existing hegemonic narrative conventions that provide what Fredric Jameson calls an ‘imaginary resolution of a real contradiction’ (62). The article concludes that the LLM presents its adaptational activity as an impartial synthesis of data qua ostensibly objective facts, obfuscating the reality that its adaptational activity is instead an expression of a political unconscious serving the interests of the ruling class. LLMs do not therefore articulate heteroglossia but rather regurgitate the biases and values of dominant ideology.