대규모 언어모델의 인간 사회 편향 학습과 재생산: 챗GPT와 딥시크는 ‘중립적 도구’인가, ‘편향적 행위자’인가 = Learning and Reproducing Human Social Biases in Large Language Models: Are ChatGPT and DeepSeek Neutral Tools or Biased Actors?

국문 초록 (Abstract)

본 연구는 대규모 언어모델(large language model [LLM])이 인간 사회의 편향을 어떻게 학습하고 재생산하는지를 실증적으로 분석하였다. 서구권 대표 모델인 챗GPT와 동양권 모델 딥시크를 대상으로 명시적(explicit) 및 암묵적(implicit) 편향의 발현 양상을 비교하였다. 인종 관련 문장을 활용한 명시적 편향 측정에서 딥시크는 전반적으로 더 엄격한 기준을 적용했으며, 챗GPT는 중립적이지만 응답 일관성이 낮았다. 또한 암묵적 연상 검사(implicit association test [IAT]) 결과, 챗GPT는 서구 개념을, 딥시크는 동양 개념을 긍정적 속성과 더 자주 연결하며 무의식적 편향을 드러냈다. 이는 LLM이 표면적으로는 중립성을 보이더라도 특정 가치체계를 강화할 수 있음을 보여주는 ‘편향의 이중성’을 시사한다. 결과적으로 LLM은 단순한 정보 생성 도구가 아니라, 학습 데이터와 설계 철학에 따라 세계관을 재현하는 사회문화적 행위자임을 확인하였다. 본 연구는 AI 윤리와 설계 과정에서 편향 인식과 조정의 필요성을 강조하며, 기술적·정책적 논의를 위한 기초 자료를 제공한다.

다국어 초록 (Multilingual Abstract)

This study examines how large language models (LLMs) such as ChatGPT and DeepSeek learn and reproduce human social biases, focusing on the differences between explicit and implicit bias. While LLMs are often framed as neutral computational tools, their outputs are shaped by their training data, prompting an examination of their role as sociocultural agents that not only reflect but also amplify existing social stereotypes. To address this issue, we designed two complementary empirical studies. The first study measured explicit bias by analyzing the models’ generated associations between racial categories and evaluative attributes, such as positive and negative descriptors. The second study employed a modified Implicit Association Test (IAT) to capture implicit bias, examining the strength of associative links between Western and Eastern concepts and various thematic categories, including leadership, competition, care, trust, failure, and innovation. A log-odds ratio (LOR) analysis was applied to quantify the relative intensity of these associations, offering a more nuanced understanding than simple frequency counts. The findings reveal striking divergences between the two LLMs. ChatGPT-4o consistently displayed stronger positive associations with Western concepts, linking them to ambition, leadership, technological progress, and problem-solving capacity, while Eastern concepts were more often tied to social harmony, care, and, at times, weakness or failure. In contrast, DeepSeek-V3 showed the reverse pattern: it reinforced positive associations with Eastern concepts, associating them with leadership, competition, and innovation, and tended to frame Western concepts less favorably. These results highlight the dual nature of bias in LLMs: explicit outputs may appear balanced or neutral, while implicit associative patterns reveal deeper cultural asymmetries embedded in training data. Such duality demonstrates that LLMs do not merely mirror reality, but actively reproduce and restructure cultural meanings in ways that align with their sociotechnical environments. By empirically comparing LLMs developed in Western and Eastern contexts, this research contributes to theoretical debates on algorithmic fairness, critical posthumanism, and the ontology of human–AI relations. It demonstrates that biases are not incidental errors but constitutive features of how LLMs process and generate knowledge, raising ethical concerns about their deployment in socially sensitive domains. The study calls for a shift from viewing LLMs as passive tools to recognizing them as active sociocultural agents, whose embedded biases require continuous critical scrutiny. In doing so, it provides a foundation for more reflexive governance and ethical design of AI systems in global contexts. The study provides meaningful insights for policymakers, technologists, and educators striving to cultivate more inclusive and accountable AI ecosystems.