김승섭 교수 “주류 기득권 관점 학습해 온 AI…사회적 불평등 심화할 수 있다”

Professor Kim Seup-sup, who studies how artificial intelligence (AI) eliminates social and minority distress, warns that large language models (LLM) absorb the hierarchy and biases inherent in the language they learn. He stresses the need to clearly recognize the limitations of current LLMs. While AI excels at operations and engineering problem-solving, it can exacerbate inequality in social issues.

Professor Kim explains that LLMs learn from the data it has absorbed and that perspective of the dominant groups is inherent in the data collected over centuries. They inherit this massive bias. Eliminating this bias, like removing a virus, may be beyond humans. Some big tech companies have tried to prevent their models from discriminating against minorities, but the problem goes beyond that.

For example, in 2024 Nature, it was reported that LLMs are positive about African Americans, but when presented with their commonly used English phrases without context, the AI judges them as lazy or stupid. The bias may not be corrected but hidden, and discrimination becomes more covert. When only a few people are represented, it significantly increases the possibility of structural discrimination.

LLMs generate many responses based on the assumption of a society without minorities, which greatly increases the risk for users. These responses can influence users. Professor Kim stresses that the question is not whether to stop, but what kind of norms to use. He believes AI will not immediately threaten humanity but will continue to exacerbate inequality.

Therefore, it is crucial to understand the risk we face and how to mitigate it. He also believes that in government and public sectors managing limited resources and representing certain populations, these AI models should not be used as they are. The future that AI will draw should not be left solely to engineers and developers. Everyone must think about what kind of norms AI will create the future.