AI는 누구의 언어로 말하는가: 포스트식민 체제하 리터러시

초록

This article begins with the question, “In whose language does AI speak?” Here, “whose” refers not to a specific subject, but rather to a planetary assemblages of technology, society, and language through which AI technology is constructing a postcolonial system by redistributing inequalities of power and resources. Generative AI operates on the basis of U.S.- and China-centered technological hegemony, data labor in the Global South, and the exploitation of natural resources rooted in geographical hierarchies; yet, by reproducing these dynamics through English-centric training data (60–90%), it perpetuates a Western-centric cultural bias (“WEIRD”). Operating within this biased value chain, generative AI’s language output automates processes such as translation, summarization, explanation, and interpretation, thereby influencing not only the existing linguistic order but also culture, cognition, and value systems. The domain where the hallucinations and biases of generative AI are most acutely revealed is “literary studies.” This study analyzed the issues of hallucinations and biases in generative AI, focusing on Korean literature—a literary text in which semantic gaps and affect play a central role and which is difficult to capture with existing language data—in addition to the non-Western language “Korean.” In particular, we examined hallucinations by distinguishing between factual hallucinations and interpretive hallucinations. While factual hallucinations are errors identified at the level of truth and falsehood—where the AI generates content that is inconsistent with or contradicts the original text as if it were factual information—interpretive hallucinations operate at the level of meaning, rearranging the text’s meaning into a probabilistically dominant semantic system while stripping away its historical, cultural, and linguistic specificity. In Korean literature—a non-Western text—this interpretive hallucination is further amplified when the Western-biased perspectives and values inherent in generative AI intervene in the interpretation of works. In short, this paper analyzed hallucinations and biases not as mere technical errors, but as symptoms of reality revealing what has been suppressed in the process of language automation. Since this symptom is not a random defect but a manifestation of the hierarchy embedded in the value chain, generative AI reinforces the postcolonial system as a symbolic linguistic system that deepens discriminatory hierarchies in technology, society, and language. This is why counter-practices that reconfigure language as a commons and critical literacy regarding Korean literature are necessary.

키워드