Presented by: Sanghoun Song
Abstract: Despite the fabulous performance of recent large language models (LLMs), two major criticisms persist. On the one hand, models such as GPT are often described as “stochastic parrots”. Stochastic parrots refer to AI systems that generate superficially human-like texts through statistical prediction without genuine comprehension. This metaphor highlights their limited cognitive plausibility and lack of meta-understanding of human language and society. On the other hand, many LLMs remain heavily reliant on English data during training, which raises doubts about their multilingual capability and cultural adaptability. In fact, these two criticisms are closely related to each other. This talk examines how these limitations are intertwined, focusing on the failure in producing contextually appropriate Korean expressions.